Fact-checked by Grok 2 weeks ago

Perceptual hashing

Perceptual hashing is a class of algorithms that generate compact, content-adaptive fingerprints for data, such as images, audio, and video, designed to produce similar values for perceptually equivalent content despite alterations like , resizing, or minor edits that do not affect human perception. These hashes prioritize perceptual invariance over exact bit-for-bit matching, enabling efficient similarity detection through metrics like , where low distances indicate near-identical perceptual features. In contrast to cryptographic hashing functions, which avalanche under even trivial input changes to ensure security and uniqueness, perceptual hashes extract robust features—often from low-frequency components via discrete cosine transforms, pixel gradients, or average intensities—to tolerate transformations while maintaining distinctiveness for dissimilar content and resilience to noise or cropping. Key implementations include average hashing (aHash), which thresholds pixel averages; difference hashing (dHash), based on adjacent pixel comparisons; and perceptual hashing (pHash), employing DCT for frequency-domain analysis, with real-world variants like Microsoft's PhotoDNA and Facebook's PDQ enhancing scalability for massive databases. The concept emerged in the early 21st century amid advances in content-based retrieval and , building on foundational hashing ideas from the mid-20th century but tailored for forensics. Notable achievements include enabling proactive detection of known abusive material, such as child sexual abuse imagery (), without requiring full-file storage, as deployed by platforms like and since around 2009. Applications span copyright enforcement, duplicate image search, tamper detection, and online , but controversies arise from trade-offs in accuracy—such as vulnerability to adversarial manipulations that preserve hashes while altering content—and privacy risks in implementations, exemplified by Apple's 2021 NeuralHash proposal, which faced scrutiny for potential false matches and enablement of broad despite its perceptual focus on hashes. Ongoing research addresses these via enhancements for better robustness, though empirical evaluations highlight persistent challenges in balancing with perceptual fidelity across diverse media.

Definition and Principles

Core Concepts

Perceptual hashing algorithms generate compact, fixed-length digital fingerprints of multimedia content, such as images, that reflect its perceptual characteristics rather than its precise binary data. These fingerprints ensure that visually or audibly similar inputs produce hash values with a measurable degree of resemblance, enabling the detection of content duplicates or near-duplicates without requiring exact matches. The core objective is to abstract invariant features of human perception, allowing hashes to serve as robust identifiers in large-scale content databases. Robustness to content-preserving modifications constitutes a primary , whereby hashes tolerate alterations like , resizing, , cropping, or low-amplitude noise that do not substantially affect perceived essence. For example, under compression at quality levels as low as 50%, effective perceptual hashes maintain similarity scores indicative of unchanged visual structure. This property arises from focusing on low-level perceptual cues, such as patterns or edge distributions, which remain stable across such transformations. Preprocessing steps, including resizing to uniform dimensions (e.g., 32×32 or 8×8 pixels) and conversion, standardize inputs to emphasize structural over chromatic details. Feature extraction underpins hash generation by isolating perceptually salient elements, often through transforms that prioritize coarse or mid-level information. (DCT) applied to low-frequency coefficients captures global and , while computations between adjacent pixels highlight local discontinuities akin to edges perceived by the human visual system. Extracted coefficients or statistics are quantized via thresholding (e.g., comparing to a value) to yield strings, typically 64 bits long, balancing compactness with discriminative power. These processes ensure even distribution of values across possible outputs, minimizing clustering and supporting efficient indexing. Similarity evaluation relies on distance metrics that quantify hash divergence in a manner aligned with perceptual tolerance. The , measuring bit mismatches as a fraction of total bits, serves as the standard; normalized values below thresholds like 0.04 or 0.3 (depending on application) denote matches, as validated in benchmarks against manipulated datasets. This approach enables probabilistic matching, where intra-class distances (similar content) remain low even after manipulations, while inter-class distances (distinct content) stay high, facilitating false positive minimization. For instance, DCT-derived hashes exhibit mean normalized s under 0.05 for additions up to standard deviation 0.01.

Distinctions from Cryptographic Hashing

Perceptual hashing functions are engineered to yield similar hash values for inputs that exhibit perceptual similarity, such as content altered by , resizing, or minor , thereby enabling robust content identification despite non-malicious transformations. In contrast, cryptographic hashing functions, such as SHA-256, rely on the , where even a single-bit change in the input produces a substantially different output, ensuring sensitivity to any alteration for applications demanding exact . This fundamental behavioral divergence stems from perceptual hashes extracting invariant features from perceptual domains—like frequency components in images—while cryptographic hashes process raw bits uniformly to prioritize unpredictability and diffusion. The purposes of these hashing paradigms further underscore their distinctions: perceptual hashes facilitate similarity matching via metrics like between fingerprints, supporting tasks such as duplicate detection and content fingerprinting in large databases, where exact matches are neither feasible nor desirable. Cryptographic hashes, however, enforce exact equality for verification, underpinning protocols including digital signatures and password storage, with properties like preimage resistance (infeasibility of reversing the to original input) and strong (computational hardness of finding distinct inputs with identical outputs). Perceptual hashes deliberately tolerate a degree of controlled collisions for perceptually equivalent content, rendering them unsuitable for cryptographic but effective for tolerant of format-preserving operations. Security trade-offs highlight additional contrasts, as perceptual hashes trade cryptographic guarantees for perceptual robustness, making them vulnerable to second-preimage attacks—where an adversary crafts a perceptually dissimilar input matching a target —or evasion by targeted perturbations that alter the hash without substantially changing . For instance, while cryptographic hashes resist by design, perceptual variants can be inverted or approximated more readily if their feature extraction is known, though this is often mitigated in practice by algorithmic secrecy or hybrid deployments. Thus, perceptual hashing prioritizes detection efficacy over adversarial hardness, inverting the evasion-forgery balance typical of cryptographic systems.

Historical Development

Origins in Content-Based Retrieval

Perceptual hashing originated from the challenges faced in content-based image retrieval (CBIR) systems during the mid-1990s, as digital image databases expanded beyond the capabilities of exact-match searches. Traditional text-based retrieval proved inadequate for visual content, prompting the development of methods to query and retrieve images based on perceptual similarity in features such as color, , and . Early CBIR systems, like IBM's Query By Image Content (QBIC) introduced in 1995, extracted low-level features from images and computed similarity using metrics like on feature vectors, enabling queries on large collections but requiring computational efficiency for scalability. The limitations of high-dimensional feature vectors—such as storage overhead and slow distance computations—drove research toward compact, robust representations that could approximate perceptual similarity while supporting fast indexing and comparison. These representations needed to tolerate minor variations like , cropping, or , mirroring human rather than bitwise exactness. In CBIR contexts, such signatures facilitated duplicate detection and near-match retrieval, forming the conceptual foundation for perceptual hashing. A pivotal advancement came in 2000 with the introduction of robust hashing by Venkatesan et al., who proposed an indexing technique using randomized on image statistics, such as wavelet coefficients, to generate hashes resilient to common distortions while resisting collisions for security. This work, motivated by content identification in retrieval scenarios, marked an early formalization of perceptual hashes as binary strings amenable to for similarity measurement, bridging CBIR's feature-based approaches with hash-like efficiency. Subsequent refinements built on these ideas, adapting them for broader retrieval tasks.

Emergence of Robust Algorithms

The limitations of early content-based retrieval systems, which relied on exact or near-exact matching and faltered under common image processing operations like or resizing, prompted the of hashing algorithms explicitly designed for perceptual robustness. In 2000, Ramarathnam Venkatesan and colleagues at introduced a pioneering robust image hashing method at the International Conference on Image Processing, utilizing randomized projections on coefficients to produce fixed-length binary sequences. This technique generated hashes resilient to manipulations such as at quality factors down to 50%, addition, and minor cropping, with empirical tests demonstrating Hamming distances under 10% for altered versions of the same image while exceeding 50% for distinct images. The ensured against preimage attacks, marking a foundational shift toward hashes that prioritized human-perceived similarity over bit-level fidelity. Building on this framework, subsequent algorithms in the early incorporated frequency-domain features to enhance invariance. For instance, methods leveraging the (DCT) low-frequency coefficients emerged around 2002–2003, extracting perceptual fingerprints by quantizing dominant DCT blocks after block-wise processing, which proved effective against rotation, scaling, and brightness adjustments in controlled experiments. These approaches achieved robustness metrics where hash collisions for perceptually similar images occurred in under 5% of cases across standard datasets like USC-SIPI, while rejecting tampered content with high specificity. The emergence of such techniques was driven by practical demands in multimedia authentication and , where cryptographic hashes failed due to their on any change, thus establishing perceptual hashing as a distinct by the mid-.

Modern Proprietary and Open-Source Advances

Microsoft's , a perceptual hashing technology first deployed in 2009 and continuously refined, normalizes images through geometric transformations and extracts features insensitive to compression or cropping, enabling platforms to match known with over 99% accuracy in controlled tests while resisting common edits. Apple's , introduced in 2021 as part of a proposed scanning system for , uses a ResNet-50 trained on diverse image datasets to generate 96-bit hashes capturing high-level semantic features, though subsequent analyses revealed vulnerabilities to black-box collision attacks allowing hash forgery with minimal perturbations. Meta's video hashing extensions, benchmarked in 2024 studies, outperform earlier image-only methods by incorporating temporal , achieving superior robustness in detecting modified clips on social platforms. Open-source libraries have advanced accessibility and customization. The pHash library, licensed under GPLv3 since its inception around with updates through the 2020s, implements DCT-based image hashing alongside radial variance for audio and block-based methods for video, supporting real-time applications like torrent monitoring for copyrighted material. Python's imagehash module, available on since 2013 and actively maintained, provides implementations of average (aHash), difference (dHash), and wavelet perceptual hashing, with thresholds tunable for duplicate detection in datasets exceeding millions of images. Meta's algorithm, developed internally from 2015 and open-sourced by 2019, employs discrete cosine transforms on perceptually weighted coefficients to yield compact 256-bit hashes, facilitating efficient nearest-neighbor searches in large-scale databases. Deep learning integrations represent cutting-edge progress. DINOHash, an open-source framework released in recent years, derives hashes from self-supervised DINOv2 embeddings, demonstrating resilience to adversarial perturbations and synthetic image alterations in verification tasks. Evaluations from 2024 highlight that such neural approaches, while improving discriminability over traditional frequency-domain methods, remain susceptible to inversion attacks reconstructing originals from hashes, prompting hybrid defenses combining hashing with . Benchmarks across , , and NeuralHash underscore trade-offs: proprietary systems excel in deployment scale but face inversion risks, whereas open-source variants enable reproducible security audits amid evolving threats like AI-generated content.

Key Algorithms and Techniques

Frequency-Domain Methods

Frequency-domain methods in perceptual hashing apply orthogonal transforms to convert data—typically images, audio, or video—into representations, emphasizing low- components that preserve essential perceptual structure while attenuating sensitivity to localized changes such as , , or minor filtering. This approach draws on the human sensory system's prioritization of low- information for overall content , enabling hashes that maintain similarity for visually or auditorily equivalent variants but diverge for substantive alterations. The (DCT) dominates image hashing implementations due to its superior energy compaction, concentrating signal power in fewer low-frequency coefficients compared to alternatives like the , which aligns with perceptual irrelevance models in compression standards such as . In a typical DCT , the input is grayscale-converted and resized to a uniform dimension (e.g., 32×32 for pHash or 64×64 for ), followed by 2D DCT application; an 8×8 or 16×16 low-frequency submatrix is then isolated, with bits derived via mean subtraction or quantization to produce 64- or 256-bit hashes, respectively. These hashes exhibit robustness to operations like resizing, blurring, or compression at quality factors above 70, though they remain vulnerable to targeted adversarial perturbations exploiting DCT's . Variants augment DCT with spatial preprocessing or for enhanced discrimination. Block-DCT schemes partition images into blocks, extract DCT coefficients alongside color histograms, apply (PCA) to fuse and compress features, and threshold for binary hashing, yielding improved tamper localization and resilience to content-preserving edits as demonstrated in 2010 experiments. Fourier-domain techniques, including the (DFT) and its derivatives like the Fourier-Mellin Transform (FMT), target rotation-scale-translation invariance by operating on log-polar representations or overlapping blocks, securing hashes with dual keys and outperforming DCT in geometric attack scenarios per 2013 benchmarks. The (DWT), providing multi-resolution decomposition, extracts approximation coefficients from frequency subbands—often in 3D for video frames— to balance robustness against rotation or cropping with computational tractability.

Spatial-Domain Methods

Spatial-domain methods for perceptual hashing process images directly in their pixel-based representation, extracting features from values, local differences, or statistical aggregates without frequency transformations such as DCT or wavelets. These approaches prioritize computational simplicity and speed, making them suitable for applications, though they often exhibit reduced robustness to geometric distortions like or cropping compared to frequency-domain counterparts. A prominent example is average hashing (aHash), which resizes the input to an 8x8 , computes the intensity across all 64 values, and generates a 64-bit binary by setting each bit to 1 if the corresponding exceeds the or 0 otherwise. This method captures global distribution but remains vulnerable to uniform adjustments, as they can flip multiple bits without altering perceptual content. Introduced as a in perceptual hashing libraries, aHash achieves high efficiency, with hashing times under 1 ms on standard hardware for typical images. Difference hashing (dHash) addresses some limitations of by emphasizing local gradients: the is resized to a 9x8 (or 8x9 for vertical variant) array, and bits are derived by comparing each to its , assigning 1 if the left is brighter or 0 otherwise, yielding a 64-bit insensitive to absolute intensity shifts. This edge-detection-like mechanism enhances discriminability for structural changes while maintaining low complexity, often outperforming aHash in stability under minor noise or compression, with inter-variant distances typically below 10 bits for perceptually similar images. Both aHash and dHash, as evaluated in comparative benchmarks, demonstrate superior speed—processing rates exceeding 1000 images per second on consumer CPUs—but trade off robustness, showing higher false negatives (up to 20-30% more under ) relative to frequency methods in standardized tests like those using Stirmark benchmarks. Advanced spatial variants, such as those incorporating block-wise statistics or cyclic coding for invariance, build on these by partitioning images into subregions and encoding relative variances, though they increase bit length to 128 or more for improved .

Neural and Learning-Based Approaches

Neural and learning-based approaches to perceptual hashing employ deep neural networks, primarily convolutional neural networks (CNNs), to automatically derive feature representations that align with human , surpassing the limitations of hand-crafted features in traditional methods by learning hierarchical invariances to manipulations like , , and compression. These systems typically involve an encoder network that maps input content to a compact , followed by a hashing module that binarizes the representation—often via thresholding or sign activation—to yield fixed-length codes, with training optimizing objectives such as contrastive loss to cluster similar perceptual instances while separating dissimilar ones. Supervised variants use labeled pairs or triplets from datasets like or custom perceptual similarity corpora, minimizing intra-class Hamming distances below thresholds (e.g., 32/256 bits) and maximizing inter-class distances. Apple's NeuralHash, released in August 2021 as part of a proposed scanning mechanism for detecting material, exemplifies this paradigm: it processes 512x512 RGB images through a modified ResNet-50 backbone with 10 residual blocks, projecting to a 256-dimensional vector before hashing via learned projections and clipping to {-1, 0, 1} values, remapped to . Trained on over a billion images with augmentations simulating device variations, it claims robustness to compression up to 70% quality loss and scaling by factors of 0.5–2.0, achieving near-zero false positives in controlled tests. However, empirical evaluations reveal critical flaws, including leakage risks and susceptibility to gradient-based adversarial perturbations that induce hash collisions with perceptual changes under 1% PSNR degradation, as demonstrated by attacks inverting hashes or dodging detection in under 100 iterations. Alternative architectures include multitask neural networks that jointly optimize perceptual hashing with tasks like autoencoding or , as in a 2021 scheme using a encoder-decoder pair trained on MSRA-B to yield 128-bit hashes resilient to (σ=0.01) and , reporting 98.5% accuracy versus 92% for DCT-based baselines. A 2022 variant introduces "hash centers" by aggregating features around image centroids post-convolution, enhancing geometric invariance for authentication; evaluated on CASIA v2.0, it maintains Hamming distances under 0.1 for tampered copies while exceeding 0.4 for forgeries, outperforming wavelet-domain methods by 15% in ROC-AUC. extensions leverage variational autoencoders or generative adversarial networks to enforce hash code orthogonality without labels, though they trade some discriminability for reduced training data needs. For video hashing, extensions incorporate temporal modeling via 3D CNNs or LSTM layers on frame sequences, capturing motion-based perceptual cues; a 2023 review notes these achieve 5–10% higher recall in duplicate detection on datasets like UCF-101 compared to 2D-only projections. Overall, these methods demonstrate superior empirical performance on metrics like under Stirmark distortions but incur higher (e.g., 10–50 ms per image on GPUs) and risks from model inversion attacks, necessitating hybrid defenses like ensemble hashing or post-hoc robustness checks.

Applications

Digital Rights Management

Perceptual hashing facilitates (DRM) by generating content fingerprints that remain consistent despite common manipulations like , resizing, or format conversion, enabling the detection of unauthorized copies of protected such as images and videos. Unlike cryptographic hashes, which detect any alteration, perceptual variants prioritize human-perceived similarity, allowing rights holders to identify infringing material with high discriminability while tolerating benign transformations. This approach underpins enforcement systems where exact matches are impractical due to inevitable signal degradations in distribution channels. In practice, perceptual hashing integrates with watermarking and technologies to create verifiable chains for digital assets. For instance, robust hash functions extract features from the (DCT) domain to embed or verify invisible watermarks, ensuring tamper detection and ownership assertion even after adversarial edits. -augmented schemes use perceptual hashes to compute similarity scores against registered originals, triggering automated licensing or takedown actions in decentralized platforms. Such systems have been proposed for video content, where (CNN)-derived hashes achieve over 95% accuracy in copy detection under , , and noise perturbations. These methods address issues in large-scale searches, outperforming traditional watermarking alone by avoiding exhaustive pixel-level comparisons. Empirical evaluations highlight perceptual hashing's efficacy in real-world scenarios, including forensic analysis of pirated media. Deep learning-based variants, such as those employing graph-embedded structures, enable coarse-to-fine retrieval of infringed 3D assets or neural models, with Hamming distances below 10% for perceptually identical copies. However, deployment requires balancing robustness against evasion risks, as minimal visual alterations can inflate distances, necessitating defenses like multi- ensembles. Peer-reviewed implementations demonstrate false positive rates under 1% for , supporting its adoption in systems for and legal .

Content Moderation and Forensics

Perceptual hashing facilitates content moderation on online platforms by generating robust fingerprints of multimedia that withstand modifications such as resizing, compression, or minor edits, enabling automated detection of known prohibited content like child sexual abuse material (CSAM). This approach compares query hashes against large databases of flagged material using metrics like Hamming distance, allowing proactive scanning of uploads without relying on exact cryptographic matches. Microsoft's , a perceptual hashing system launched in 2009 through collaboration with , is a primary tool for detection; it creates irreversible image signatures resilient to perceptual changes and has been provided free to the National Center for Missing & Exploited Children (NCMEC) and since its donation, with cloud access via starting in 2015. Adopted by major tech firms and nonprofits, has supported the identification of millions of exploitation instances by matching variants of confirmed illegal images. Open-source libraries like pHash similarly underpin filtering systems for inappropriate visuals in user-generated content. In , perceptual hashing supports by enabling approximate matching of manipulated evidence, such as altered images in investigations, where exact hashes fail due to edits or formats. Tools like the framework allow forensic experts to test algorithms on bespoke datasets, optimizing discriminability for tasks including tracing dissemination in encrypted channels via targeted scanning. This method aids in and linkage across seizures, prioritizing perceptual similarity over byte-level identity.

Duplicate Detection and Retrieval

Perceptual hashing supports duplicate detection by generating compact, content-derived fingerprints that tolerate perceptual variations like , resizing, or cropping, unlike cryptographic es which demand exact matches. Systems compute a for incoming and measure its against stored hashes in a database; distances below a tuned —typically 5-10 bits for 64-bit hashes—flag potential duplicates, enabling automated filtering in photo libraries or archives. This method scales to millions of items via indexing techniques, such as custom hash tables that accelerate lookups by up to 300% over linear scans. In retrieval contexts, perceptual hashes index for content-based similarity searches, where a query hash retrieves nearest neighbors representing visually akin files. For images, (DCT)-based algorithms like pHash extract low-frequency coefficients to form rotation- and scale-invariant representations, supporting applications in and forensic analysis. Video retrieval employs frame-aggregated hashes robust to temporal edits, as in tools generating 64-bit fingerprints for near-duplicate clips under format distortions. Empirical implementations demonstrate efficacy in large datasets; for instance, perceptual hashing baselines achieve precise near-duplicate filtering when hybridized with neural networks, outperforming standalone exact matching in recall for transformed content. In , hashing integrates with or Gabor filters to enhance query precision, facilitating rapid location of similar assets without exhaustive comparisons. Such systems prioritize discriminability, with Hamming thresholds calibrated to balance false positives against computational overhead in real-time scenarios.

Evaluation and Performance Metrics

Robustness and Discriminability

Robustness in perceptual hashing denotes the stability of hash outputs against content-preserving transformations, such as compression, addition, scaling, and minor rotations, where similar inputs should yield hashes differing by few bits (typically <5-10 in 64-bit schemes). Evaluations commonly apply standardized manipulations to benchmark datasets like FVC 2000 or ImageNet subsets, measuring mean normalized s or bit error rates post-transformation. For example, under compression at quality 40, average hashing (aHash) achieves mean distances of 0.001-0.035, outperforming singular value decomposition-based hashes (SVD-Hash) which exceed 0.2, indicating superior tolerance to lossy encoding in simple spatial methods. Frequency-domain approaches like perceptual hash (pHash) excel against compression artifacts due to reliance on low-frequency discrete cosine transform coefficients, maintaining low bit flips even at aggressive quality reductions, though vulnerability increases with geometric shifts beyond 2 degrees rotation. Discriminability, conversely, assesses the hash's ability to differentiate perceptually distinct images via high inter-hash distances, minimizing false positives through low collision probabilities at operational thresholds. This is quantified using normalized Hamming distance distributions, where collision probability P_c is derived from mean and standard deviation of distances across dissimilar pairs, ideally approaching zero for thresholds around 0.04-0.08. pHash demonstrates strong performance here, with P_c \approx 0 \times 10^{-2} at threshold 0.04 on fingerprint image corpora, enabling precise retrieval while aHash prioritizes robustness at the cost of slightly elevated collisions. In authentication contexts, discriminability contributes to high precision and recall; pHash yields F1-scores of 0.905 across manipulations, reflecting balanced separation of tampered versus intact content. An inherent trade-off exists: enhancing robustness via longer hashes (e.g., 256 bits in PDQ) or smoothed features improves invariance but can degrade discriminability under adversarial perturbations, where bit error rates exceed 99% success for evasion at thresholds like 10 for pHash. Empirical tests reveal spatial methods like difference hash (dHash) favor scaling robustness (low distances post-resizing) but falter in noise-heavy scenarios compared to pHash, with overall discriminability following near-normal distance distributions for random pairs. Advanced schemes like PhotoDNA resist untargeted evasion (attack success rates <1% for PDQ equivalents) yet show 92% vulnerability in black-box settings without defenses, underscoring causal limits from linear feature approximations.
AlgorithmKey StrengthExample Metric (JPEG Q=40)Collision Prob. (T=0.04)
aHashRobustness to noise/compressionMean HD=0.001Higher (~10^{-1})
pHashBalanced discriminabilityMean HD=0.01-0.05~0 × 10^{-2}
dHashScaling invarianceMean HD=0.02Low, normal distribution
SVD-HashPoor overallMean HD>0.2Elevated
This table summarizes comparative performance from controlled experiments on 800-image sets, highlighting algorithm-specific profiles without universal superiority.

Computational Efficiency

Spatial-domain perceptual hashing algorithms, such as average hash (aHash) and difference hash (dHash), prioritize efficiency through minimal preprocessing, typically resizing images to low resolutions like 8×8 or 9×8 pixels followed by elementary operations—mean pixel value thresholding for aHash or adjacent pixel differencing for dHash—yielding hashes in constant time for fixed-size inputs and enabling sub-second processing even on legacy hardware. These methods avoid transformative computations, making them suitable for high-throughput scenarios like large-scale duplicate detection, where dHash demonstrates superior runtime to alternatives like pHash in empirical evaluations. Frequency-domain techniques, including pHash via (DCT) on 32×32 grayscale images, introduce logarithmic overhead from the transform (O(n log n) for n≈64 coefficients post-resizing), resulting in slower extraction; a 2010 benchmark across 94 images reported ~9.7 seconds per image for DCT-based hashing on an Intel Core 2 Duo processor, compared to ~0.6 seconds for block-mean-value spatial hashing akin to aHash. Wavelet-based variants (wHash) similarly elevate costs through discrete wavelet transforms but retain practicality for static images, with total runtimes scaling linearly with input complexity yet remaining under milliseconds on modern CPUs for single instances. Neural and learning-based approaches amplify demands via convolutional layers or , often requiring GPU acceleration for viability; group-wise CNN hashing mitigates per-image costs but still exceeds traditional methods by orders of magnitude in training phases, limiting deployment to server-side forensics over . In video benchmarks, perceptual systems like vPDQ achieve 0.004–0.007 seconds per video-second for hashing, versus 0.009–0.017 for , underscoring efficiency-robustness trade-offs where faster algorithms sacrifice recall in matching. Hamming distance computations for similarity, central to evaluation, add negligible overhead (O(hash length), typically 64 bits), but scalability in databases relies on approximate nearest-neighbor indexing, with spatial methods' simplicity facilitating lower storage (e.g., 8 bytes per hash) and faster queries than transform-heavy counterparts. Overall, efficiency favors spatial over frequency-domain methods for resource-constrained environments, as confirmed by runtime analyses prioritizing dHash for code computation speed.

Limitations and Technical Challenges

Inherent Trade-Offs

Perceptual hashing algorithms inherently balance robustness, the capacity to generate similar hash values for content subjected to benign modifications such as compression, resizing, or minor noise addition, against discriminability, the ability to produce dissimilar hashes for perceptually distinct to minimize false matches. This trade-off arises because perceptual similarity exists on a , yet hashing requires decisions that amplify small perceptual differences into hash collisions or misses when robustness is prioritized. Enhancing robustness, for instance by incorporating invariant features like discrete cosine transforms, often broadens the tolerance for transformations, thereby increasing the risk of conflating unrelated and elevating false positive rates. The tension manifests in threshold-based comparisons using metrics like normalized , where intra-class distances (for similar content) must remain low (e.g., below 0.05 for robust matches) while inter-class distances approach 0.5 for effective discrimination. (ROC) curves and area under the curve () quantify this balance, with higher values (e.g., approaching 1.0) indicating superior trade-offs, as seen in evaluations of algorithms like pHash, which achieve mean bit error rates under 0.1 for perturbations like 10% compression while maintaining collision probabilities near zero for distinct pairs. Parameter tuning, such as adjusting similarity , modulates false negative rates (missed detections) against false positives, but lowering thresholds to boost robustness can spike false positives exponentially—for example, a of 0.05 in pHash variants yields over 0.1% false positive rates on large-scale sets, potentially flagging millions of benign files daily. A secondary inherent involves computational versus perceptual , as more robust schemes relying on complex feature extraction (e.g., local descriptors or neural embeddings) demand higher processing overhead, limiting scalability in applications like . Shorter lengths improve and query speed but degrade discriminability by increasing random collision probabilities, approaching 50% bit error rates for unrelated only with longer, costlier representations. These compromises stem from the fuzzy nature of human , which defies perfect mathematical without domain-specific adaptations that still fail under diverse manipulations like geometric distortions beyond 20 degrees .

Empirical Shortcomings

Empirical evaluations of perceptual hashing algorithms reveal substantial limitations in robustness to common image manipulations, often resulting in elevated false negative rates () or false positive rates (FPR) that undermine practical utility. For instance, in assessments across social media platforms like , , and , the (DCT)-based pHash demonstrated an of 12.96% under manipulations including , , addition, and compression, while Marr-Hildreth edge detection-based dHash (often aligned with difference hashing variants) yielded an FPR of 35.18%, indicating frequent mismatches due to platform-specific processing artifacts such as aggressive compression. Similarly, (SVD)-based methods exhibited the highest at 38.89%, highlighting sensitivity to and that alters low-frequency components critical for hash stability. Large-scale empirical tests on datasets like (over 1.1 million images) further expose the trade-offs in threshold selection for comparisons. At stricter thresholds (e.g., pHash T=2), exceeds 97.8% for even mild benign transformations such as resizing or cropping, failing to detect perceptually similar content; loosening thresholds to T=14 inflates FPR to 73%, generating millions of erroneous matches in databases of comparable size. These shortcomings stem from the algorithms' reliance on fixed feature extractions—such as average intensity for aHash or gradient differences for dHash—which degrade under real-world variations like filtering or minor geometric shifts, producing s that cross decision boundaries unpredictably. In manipulation detection scenarios, perceptual hashes often underperform against combined alterations, with studies reporting less-than-ideal discrimination between benign edits and malicious forgeries; for example, wavelet-based variants (sometimes grouped with average hashing) achieve only moderate reductions (5.5%) but falter on rotation-heavy datasets due to phase shifts disrupting coefficient alignments. Such empirical gaps underscore the algorithms' brittleness in diverse, uncontrolled environments, where content-specific factors like texture density or color histograms amplify distance variances beyond typical thresholds (e.g., 5-10% Hamming allowance), necessitating dataset-specific tuning that limits generalizability.

Security Vulnerabilities

Evasion and Inversion Attacks

Evasion attacks on perceptual hashing involve adversarial modifications to input , such as images, that alter the resulting sufficiently to mismatch a target in a database while preserving perceptual similarity to the original content. These attacks exploit the gradual change in outputs under small perturbations, enabling illicit material to bypass detection systems like those used for . In black-box settings, where attackers lack access to the hashing model's internals, evasion remains feasible; for instance, evaluations of , , and NeuralHash demonstrate success rates exceeding 90% for targeted evasion with minimal perceptual distortion, often using gradient-free optimization or surrogate models. Such attacks typically employ techniques like adding imperceptible noise or applying content-preserving transformations (e.g., slight rotations, compressions, or color shifts) to cross the threshold for non-matching hashes. Deep learning-based perceptual hashers, including NeuralHash, prove particularly susceptible in white-box scenarios, where adversaries can compute gradients to minimize hash similarity efficiently, achieving evasion with perturbations below human detection thresholds (e.g., PSNR > 40 dB). Traditional algorithms like exhibit relative robustness to untargeted evasion but falter against targeted attacks mimicking specific database entries, as shown in large-scale experiments on datasets like . Inversion attacks, conversely, aim to reconstruct an approximation of the original input from its perceptual alone, potentially compromising by generating visually similar images that evade removal tools or reveal sensitive details. These attacks leverage the invertibility of hash functions, particularly for shorter hashes (e.g., 256 bits in ), using optimization methods to solve for inputs yielding the target hash digest. Recent assessments reveal that and succumb to low-computational-cost inversions, producing images with structural similarity indices (SSIM) above 0.7 to originals, sufficient for fooling downstream detectors in image-based material removal systems. NeuralHash shows greater resistance to inversion due to its learned embeddings, requiring higher computational budgets (e.g., thousands of iterations) for viable reconstructions, yet vulnerabilities persist in constrained hash spaces. Empirical tests indicate that inversion success correlates with hash length and steps, where shorter representations amplify collision risks during reconstruction. Overall, both attack types underscore perceptual hashing's tension between robustness to benign edits and fragility against deliberate adversaries, prompting calls for hybrid defenses like ensemble hashing or cryptographic commitments.

Collision and False Positive Risks

Perceptual hashing algorithms are susceptible to collisions, where distinct inputs yield hashes with sufficiently low Hamming distances to be classified as matches under a given , resulting in false positives that misidentify dissimilar content as identical. This risk arises from the inherent design prioritizing perceptual similarity over cryptographic uniqueness, allowing minor perceptual variations—such as artifacts, resizing, or color shifts—to preserve hash proximity while enabling unintended overlaps between unrelated media. Unlike exact-match cryptographic , perceptual variants exhibit probabilistic collision behaviors influenced by the algorithm's feature extraction (e.g., DCT coefficients in pHash or gradients in dHash), with false positive rates (FPRs) empirically varying by selection to balance robustness against discriminability. Quantitative evaluations reveal significant variability in collision risks across algorithms. For instance, Microsoft's , widely used in , achieves an estimated false positive probability of 1 in 10 billion for exact matches against known hashes, as validated in deployment-scale testing, though this assumes stringent thresholds and may degrade under heavy modifications. In contrast, open-source algorithms like pHash, aHash, and dHash show higher FPRs in large-scale benchmarks on datasets such as : at conservative thresholds (e.g., ≤2 for pHash), FPRs hover around 0.1%, but loosening to ≤14 for robustness elevates them to 73%, potentially generating millions of daily false alarms in platforms processing billions of images. maintains tighter inter-image Hamming distributions ( 0.5000, deviation 0.0321), minimizing false positives compared to less robust options like ColorHash, which forms classes exceeding 20,000 images due to color-based hashing insensitivity. These risks amplify in massive databases via probabilistic effects akin to the birthday paradox, where even FPRs below 10^{-9} can yield expected collisions in collections exceeding 10^9 items, as seen in natural overlaps within for algorithms like NeuralHash. Poorly tuned thresholds exacerbate this, with intra-variant distances (e.g., mirroring in pHash averaging 0.4904) sometimes overlapping inter-distances, leading to systemic false positives in forensics or duplicate detection. Vendor claims, such as PhotoDNA's low rates, warrant scrutiny against independent adversarial evaluations, which confirm baseline robustness but highlight sensitivity to dataset biases and transformations like borders or watermarks that artificially inflate similarity scores. Mitigation often involves hybrid thresholds or multi-algorithm ensembles, yet no perceptual hash eliminates collision risks without sacrificing utility.

Controversies and Ethical Implications

Privacy and Surveillance Debates

Perceptual hashing technologies, such as Microsoft's and Apple's NeuralHash, have been deployed by tech companies to detect known instances of material (CSAM) and other illegal content in user-uploaded media, by comparing perceptual hashes against curated databases without transmitting full images to servers. This client-side scanning (CSS) approach aims to preserve in services like or by performing matches on-device, flagging only potential matches for human review, thereby minimizing raw data exposure. Proponents, including and advocates, argue that such systems enable proactive moderation of encrypted communications while upholding privacy through hash-based similarity detection rather than content inspection, with empirical data from deployments showing millions of CSAM detections annually by platforms like and . Critics contend that perceptual hashing in CSS frameworks introduces systemic surveillance risks, as on-device scanning creates a precedent for mandatory inspections that could extend beyond CSAM to political dissent or other disfavored content via government-mandated hash database expansions, evidenced by proposals like the EU's 2022 Chat Control initiative requiring CSS in encrypted apps. Privacy advocates, including the , highlight inversion attacks where adversaries reconstruct prohibited images from leaked hash sets, potentially enabling targeted harassment or database compromises, as demonstrated in 2021 research inverting NeuralHash outputs with high fidelity using modest computational resources. False positive rates, while low in controlled tests (e.g., PhotoDNA's 1 in 50 billion for random images), amplify in diverse global datasets, risking erroneous flags on innocuous family photos and eroding user trust, particularly in jurisdictions with histories of authoritarian overreach. The 2021 Apple scanning proposal, which integrated NeuralHash into for iCloud photo checks, ignited global backlash, leading to its suspension amid concerns over and the feasibility of evading safeguards like hash blinding, with audits revealing to adversarial perturbations that preserve visual similarity while altering es. Empirical evaluations of practical perceptual hashing algorithms underscore trade-offs where robustness against evasion enhances discriminability but heightens inversion risks, complicating claims of preservation in contexts. While peer-reviewed studies affirm hashing's utility for targeted detection without wholesale decryption, systemic biases in database curation—often reliant on Western-centric inputs—raise equity issues, potentially overlooking culturally variant illegal content while over-flagging minority-group media. Debates persist on verifiable oversight mechanisms, such as hash set audits, to mitigate , though failures in prototypes suggest causal pathways from flaws to broader erosions of norms.

Implementation Failures and Backlash

Implementation failures in perceptual hashing deployments frequently manifest as elevated false positive rates due to hash collisions, where perceptually dissimilar content yields matching or near-matching hashes under operational thresholds. Large-scale evaluations of systems like , , and NeuralHash reveal that these algorithms produce unacceptably high false positives in client-side scanning scenarios, particularly when adversarial perturbations or dataset poisoning exploit the limited bit-length representations, undermining reliability in pipelines. A notable instance occurred in video perceptual hashing benchmarks, where false positives arose from uniform or low-contrast frames, such as dark scenes lacking distinctive features, leading to erroneous matches against known illicit content databases and requiring algorithmic refinements or post-processing filters. These shortcomings highlight how implementation choices, including threshold calibration and preprocessing, can amplify inherent discriminability limits, resulting in operational inefficiencies and the need for costly human oversight in production environments. Public backlash peaked with Apple's NeuralHash initiative, announced on August 5, 2021, which employed perceptual hashing for on-device detection in Photos via modified image matching against a hashed database from the National Center for Missing & Exploited Children. Researchers promptly demonstrated collisions by crafting innocuous images that hashed identically to targets, exposing the system's vulnerability to inversion attacks and questioning its claimed low rate of approximately 1 in 1 trillion per 1,000-image library. Critics, including privacy advocates and cryptographers, decried the approach for risking under the guise of safety, arguing that even rare errors at Apple's ecosystem scale could generate widespread false reports, erode integrity, and invite regulatory abuse. Sustained opposition prompted Apple to pause rollout on , 2021, and abandon the project entirely on December 7, 2022, citing insurmountable technical hurdles and societal trade-offs in versus detection efficacy. This retreat exemplified how perceptual hashing's deployment pitfalls—combining technical fragility with ethical overtones—can provoke decisive rejection, tempering enthusiasm for similar proactive scanning technologies in consumer platforms.

References

  1. [1]
    [PDF] An Overview of Perceptual Hashing - Semantic Scholar
    In the broadest use of the term, there are many applications of hashing, including efficient data storage and retrieval, data integrity, and data matching. We ...<|separator|>
  2. [2]
    A Survey of Perceptual Hashing for Multimedia - ACM Digital Library
    Jul 19, 2025 · This survey provides a systematic overview of the definition, basic steps, main characters, and application scenarios of perceptual hashing.
  3. [3]
    PhotoDNA | Microsoft
    PhotoDNA is an incredible tool to help detect, disrupt and report the distribution of child exploitation material.PhotoDNA FAQ · PhotoDNA Cloud Service · Documentation · Terms of Use
  4. [4]
    [PDF] Implementation and Benchmarking of Perceptual Image Hash ...
    Perceptual image hash functions produce hash values based on the image's visual appearance. A perceptual hash can also be referred to as e.g. a robust hash or ...Missing: principles | Show results with:principles<|separator|>
  5. [5]
    [PDF] A comparative study of perceptual hashing algorithms - CEUR-WS
    The objective of this paper is to analyze and evaluate the performance of five of the commonly used perceptual hashing techniques when applied to fingerprint ...
  6. [6]
    [PDF] Evaluating Perceptual Hashing with Machine Learning - IACR
    Nov 18, 2021 · Perceptual hashing uses functions like PhotoDNA to create a compact image digest, producing similar digests for similar-looking files, used to ...
  7. [7]
    A Secure and Robust Image Hashing Scheme Using Gaussian ... - NIH
    Compared to cryptographic hash functions, an image hash or a Perceptual Hash Function (PHF) is resilient to content preserving distortions and sensitive to ...
  8. [8]
    Perceptual versus cryptographic hashes for CSAM scanning
    Aug 24, 2021 · Perceptual hashes attempt to address this issue by trading off increased ease of forgery for decreased ease of evasion. They're designed so that ...
  9. [9]
    Query by image and video content: the QBIC system - IEEE Xplore
    We have developed the QBIC (Query by Image Content) system to explore content-based retrieval methods. QBIC allows queries on large image and video databases ...
  10. [10]
    Robust image hashing | IEEE Conference Publication - IEEE Xplore
    This paper introduces a novel image indexing technique that may be called an image hash function. The algorithm uses randomized signal processing strategies for ...
  11. [11]
    (PDF) Robust image hashing - ResearchGate
    Aug 6, 2025 · Venkatesan et al. [13] in 2000 pioneered the field by proposing the first image hashing system that employed randomized signal processing ...
  12. [12]
  13. [13]
    The Landscape of CSAM Detection: Challenges and Innovations
    Oct 16, 2023 · By far the most prominent among the CSAM image hash matching systems is PhotoDNA. Developed by Microsoft for use on Bing and SkyDrive but ...
  14. [14]
    [PDF] Black-box Collision Attacks on Apple NeuralHash and Microsoft ...
    Aug 26, 2025 · The 2010 thesis of Zauner [68, 37] provides a detailed introduction to perceptual hash function and introduces the open-source construction ...
  15. [15]
    [PDF] A Benchmark-based Analysis of Perceptual Hash Systems for Videos
    Sep 26, 2024 · Perceptual hash algorithms measure similarity between videos, used to detect CSAM by comparing content against known databases.
  16. [16]
    pHash.org: Home of pHash, the open source perceptual hash library
    pHash is an open source software library released under the GPLv3 license that implements several perceptual hashing algorithms.Demo · Apps · Docs · Support
  17. [17]
    JohannesBuchner/imagehash: A Python Perceptual Image Hashing ...
    An image hashing library written in Python. ImageHash supports: Average hashing; Perceptual hashing; Difference hashing; Wavelet hashing; HSV color hashing ( ...
  18. [18]
    None
    Nothing is retrieved...<|separator|>
  19. [19]
    [PDF] DINOHash: Learning Adversarially Robust Perceptual Hashes from ...
    We present DINOHash, an open- source framework for robust perceptual image hashing designed for the provenance detection of. AI-generated images. Unlike digital ...
  20. [20]
    Robustness of Practical Perceptual Hashing Algorithms to ... - arXiv
    Dec 5, 2024 · This paper assesses the security of three widely utilized PHAs—PhotoDNA, PDQ, and NeuralHash—against hash-evasion and hash-inversion attacks ...
  21. [21]
    [PDF] Evaluating the robustness of perceptual hashing-based client-side ...
    Designed to be robust to small changes to the media, as well as transformations like rotation and rescal- ing, perceptual hashing algorithms would be used to ...<|separator|>
  22. [22]
    A Robust and Discriminative Image Perceptual Hash Algorithm
    Image perceptual hashing find applications in content indexing, large ... We propose a Block-DCT and PCA based image perceptual hash in this ...
  23. [23]
    Fourier-Mellin transform for Robust Image Hashing
    ### Summary of Fourier-Mellin Transform Based Perceptual Image Hashing Technique
  24. [24]
    [PDF] Exploring Spatial Encoding in Perceptual Hashes | DFRWS
    Determine if redundant data in an image is identifiable in output Perceptual. Hashes. Insert constant data at fixed positions (e.g., top-left, border, ...
  25. [25]
  26. [26]
    [2108.11794] State of the Art: Image Hashing - arXiv
    Aug 26, 2021 · In this article, we present a short review of the state-of-the-art traditional perceptual hashing and deep learning-based perceptual hashing methods.
  27. [27]
    Learning to Break Deep Perceptual Hashing: The Use Case NeuralHash
    ### Summary of Deep Perceptual Hashing and NeuralHash
  28. [28]
    Perceptual Image Hashing Based on Multitask Neural Network - 2021
    Dec 18, 2021 · In the methods based on deep neural networks, the scheme in [24] proposed a robust image hashing scheme, which is based on deep learning.
  29. [29]
    Learning to Break Deep Perceptual Hashing: The Use Case ... - arXiv
    Jul 16, 2024 · The more recent deep perceptual hashing algorithms utilize neural networks for feature extraction. Perceptual hashing algorithms differ ...
  30. [30]
    Deep Perceptual Hash Based on Hash Center for Image Copyright ...
    Nov 14, 2022 · A novel Convolutional Neural Network (CNN)-based perceptual image hashing scheme is introduced in this paper.
  31. [31]
    An Image Perceptual Hashing Algorithm Based on Convolutional ...
    Jan 1, 2023 · To address this problem, an image-perceptual hashing algorithm based on convolutional neural networks is proposed in this paper. First of ...
  32. [32]
    PHASER: Perceptual hashing algorithms evaluation and results
    We present PHASER, a still-image perceptual hashing framework enabling forensics specialists and scientists to conduct experiments on bespoke datasets.Missing: core | Show results with:core
  33. [33]
    An Overview of Perceptual Hashing
    Oct 28, 2021 · Perceptual hashing has emerged as a powerful technology to limit the redistribution of multimedia content (including audio, images, and video).Missing: early papers 1990s
  34. [34]
    An Improved Design Scheme for Perceptual Hashing Based on ...
    In an earlier work, we proposed a digital rights management system based on digital watermarking, blockchain, and perceptual hashing to resolve these issues.
  35. [35]
    [PDF] Modeling of Image Copyright Protection using Discrete Cosine ...
    The perceptual hashing algorithm in this research is used the DCT hash algorithm which is the most robust perceptual hash algorithm (Drmic et al. 2017). The DCT ...
  36. [36]
    Digital rights management scheme based on redactable blockchain ...
    Sep 4, 2023 · The perceptual hash is adopted to detect similarities and judge whether or not it is the original piece. In addition, an incentive scheme based ...
  37. [37]
    A robust and secure perceptual hashing system based on a ...
    Perceptual hashing schemes are inspired from the cryptographic hash functions to authenticate multimedia. Traditionally, data integrity issues are addressed by ...
  38. [38]
    Graph-Embedded Structure-Aware Perceptual Hashing for Neural ...
    Model perceptual hashing offers an efficient solution for retrieving the pirated models. Existing methods, such as handcrafted feature-based and dual-branch ...
  39. [39]
    [PDF] It's Not What It Looks Like: Manipulating Perceptual Hashing based ...
    We select two popular perceptual hashing algorithms that are widely used by researchers and industry practitioners. • Standard pHash. pHash is a widely used ...
  40. [40]
    Perceptual Hashing based on Machine Learning for Blockchain and ...
    In this current study, to obtain a stable message digest, we propose a method of improving perceptual hashing based on machine learning. With this method, an ...
  41. [41]
    Home of pHash, the open source perceptual hash library - Apps
    BitBlocker uses the advanced perceptual hashing algorithms in pHash to actively scan file sharing sites and BitTorrent trackers for copyrighted material ...Missing: modern | Show results with:modern
  42. [42]
    [PDF] Using Perceptual Hashing for Targeted Content Scanning
    A perceptual-hashing-based targeted content scanning (PHTCS) system com- prises three core components, a perceptual hash function, the targeted content.
  43. [43]
    [PDF] Hamming distributions of popular perceptual hashing techniques
    Perceptual Image Matching. Perceptual hashing (Farid, 2021; Hadmi et al., 2012) approaches, which are inspired by the domain of Content-Based Image Retrieval.<|separator|>
  44. [44]
    akamhy/videohash: Near Duplicate Video Detection ... - GitHub
    Videohash is a Python package for detecting near-duplicate videos (Perceptual Video Hashing). It can take any input video and generate a 64-bit equivalent hash ...
  45. [45]
    Effective near-duplicate image detection using perceptual hashing ...
    May 27, 2025 · Perceptual hashing gives us a quick way to filter out similar-looking pictures, while the Siamese network architecture paired with the Vision ...
  46. [46]
    Perceptual Hashing for Content Based image Retrieval - IEEE Xplore
    In this paper, CBIR using Perceptual hashing is implemented with feature extraction techniques like Color histogram, Gabor filters, and Canny's edge detection.Missing: origins | Show results with:origins
  47. [47]
    Home of pHash, the open source perceptual hash library
    pHash is capable of computing hashes for audio, video and image files. You start by computing hash values for the media files.
  48. [48]
    None
    ### Summary of Robustness of PhotoDNA, PDQ, and NeuralHash to Manipulations
  49. [49]
    Distance distributions and runtime analysis of perceptual hashing ...
    Perceptual image hashing refers to a class of algorithms that produce content-based image hashes. These systems use specialized perceptual hash algorithms ...
  50. [50]
  51. [51]
    Perceptual Image Hashing Fusing Zernike Moments and Saliency ...
    This paper proposes a novel perceptual image hashing scheme that robustly combines global structural features with local texture information for image ...<|separator|>
  52. [52]
    [PDF] Evaluation of Perceptual Hashing Algorithms Against Image ...
    In defending against image-based fake news, compared with crypto hash, perceptual hash shows its advantage in detecting image manipulation while still being ...
  53. [53]
    Robustness of Practical Perceptual Hashing Algorithms to ... - arXiv
    Jun 3, 2024 · This paper assesses the security of three widely utilized PHAs - PhotoDNA, PDQ, and NeuralHash - against hash-evasion and hash-inversion attacks.
  54. [54]
    Evaluating the robustness of perceptual hashing-based client-side ...
    In a large-scale evaluation, we show perceptual hashing-based client-side scanning mechanisms to be highly vulnerable to detection avoidance attacks in a black- ...
  55. [55]
  56. [56]
    [PDF] Hamming Distributions of Popular Perceptual Hashing Techniques
    Dec 16, 2022 · The chosen hash algorithms are listed in Table 1. The list covers: frequency transforms; colour his- tograms; block mean colour; and deep ...
  57. [57]
    [PDF] PhotoDNA - ITU
    likelihood of false positive of 1 in 10 billion. Once NCMEC assigns PhotoDNA signatures to known images of abuse, those signatures can be shared with online ...
  58. [58]
    Towards arbitrating in a dispute - on responsible usage of client-side ...
    Jun 5, 2024 · CSS is typically based on perceptual hashing, hence we call this approach Perceptual Hash-based Targeted Content Scanning (PHTCS). This paper ...
  59. [59]
    The Ethical and Social Implications of Perceptual Hashing
    Jun 12, 2024 · The debate around perceptual hashing is one such issue that demands greater attention. ... perceptual hashing for surveillance and censorship.
  60. [60]
    [PDF] Public Verification for Private Hash Matching
    As we discuss further below, an untrustworthy hash set or an untrustworthy implementation creates serious risks for security, privacy, and free expres- sion ...
  61. [61]
    [PDF] Perceptual Hash Based Content Matching
    Chapter VII discusses the use of perceptual hash based content matching for AI generated content. Finally, Chapter. VIII concludes the thesis by summarizing the ...
  62. [62]
    Apple says collision in child-abuse hashing system is not a concern
    Aug 18, 2021 · Researchers have produced a collision in iOS's built-in hash function, raising new concerns about Apple's CSAM-scanning system.
  63. [63]
    NeuralHash Collisions: The Limits of Apple's Image Hashing
    Aug 19, 2021 · Apple's NeuralHash: real-world collisions, false positives, and implications for CSAM detection. See what this means for system reliability.
  64. [64]
    Apple Backs Down on Its Controversial Photo-Scanning Plans
    Sep 3, 2021 · A sustained backlash against a new system to look for child sexual abuse materials on user devices has led the company to hit pause.Missing: discontinues | Show results with:discontinues
  65. [65]
    Apple abandons controversial plan to check iOS devices and iCloud ...
    Dec 8, 2022 · Apple is abandoning its plans to launch a controversial tool that would check iPhones, iPads and iCloud photos for child sexual abuse material (CSAM)Missing: discontinues | Show results with:discontinues