Western blots are images of protein bands on a membrane, and they are the most common target of inappropriate image reuse in biomedical papers. They are hard to check because a panel is little more than dark horizontal smudges on a grey rectangle. The check that would matter most compares a newly submitted blot against unpublished raw scans held by other labs, journals and companies. Nobody will pool those scans in the clear. Secure deduplication, the storage-systems practice of spotting identical data across users without the server reading it, looks like the natural tool. It fails in two ways. It only recognises bit-identical files, and the fuzzy perceptual hashes used to extend it to images leak content: a server can guess low-entropy images and confirm its guesses against the stored hashes.
We propose fingerprinting a blot by the random microtexture of its membrane background (speckle, grain, dust and bubble marks that are unique to one physical exposure) rather than by its band pattern. The fingerprint is then passed through a keyed, rate-limited hashing service before any shared server sees it. The band pattern is exactly what an attacker can guess and what honest, independent blots share. A β-actin loading-control band at 42 kDa looks the same in thousands of papers. The membrane texture is what makes two crops "the same piece of film." Moving the fingerprint onto the texture should therefore buy two things together: fewer false alarms on legitimately similar blots, and resistance to the guessing attacks that break perceptual-hash deduplication.
Core Claim
Background-texture fingerprints should beat band-content fingerprints on both axes that matter here. On specificity, they should separate "same membrane, re-cropped or contrast-adjusted" from "different membrane, same protein." On privacy, the guesser's uncertainty about the fingerprint (its min-entropy) should be much higher, because the speckle of an unpublished scan cannot be reconstructed from knowledge of the biology. That property is what lets a cheap keyed-hash protocol stand in for heavy cryptography. The claim can be falsified directly. If texture does not survive journal-style processing (JPEG, resizing, contrast changes), or if band-content tags under the same keyed protocol match it, the idea fails.
Positioning Against Prior Work
Three plaintext forensics systems motivate the problem and supply parts of the method. Learning to identify image manipulations in scientific publications extracts Western blot regions with rectangle heuristics and classifies pairs with a twin CNN. It reports 90.86% accuracy on the Bik duplication data and about a 13% gain from region isolation. We reuse its region step. Image Forensics: Detecting duplication of scientific images with manipulation-invariant image similarity learns a 128-dimensional triplet embedding with a distance threshold of 1.0, and it is known to drift across that threshold under strong histogram and contrast changes. This weakness is part of why we normalise intensity by rank. Forensic Analysis of Synthetically Generated Western Blot Images shows that texture co-occurrence statistics separate real blots from generated ones, which supports our premise that background texture carries physical-origin information. It also reports that resizing hurts those statistics, which is a constraint we must test head-on. All three systems operate on unencrypted images, and none addresses cross-institution privacy.
On the deduplication side, Image Deduplication Based on Hashing and Clustering in Cloud Storage is the closest system. It sends a texture-plus-frequency perceptual hash in the clear next to the ciphertext and matches by Hamming distance. We differ in two ways: we never expose the hash, and we fingerprint background rather than whole-image content. SimEnc maps similar Docker chunks to shared keys through a learned hash and clustering. It admits that this leaks similarity relationships and that its network caps throughput, which is a warning for our design. Secure Deduplication of Encrypted Data: Refined Model and New Constructions formalises offline brute-force attacks on low-entropy content. That analysis is the threat model our entropy argument answers. Information Leakage in Encrypted Deduplication via Frequency Analysis warns that match counts themselves leak information, which bounds what any tag scheme can promise. Heavier primitives show the privacy ceiling and its cost: attribute-based equality tests over pairings in Attribute-Based Equality Test Over Encrypted Data Without Random Oracles, and zero-knowledge ownership proofs in Private data deduplication protocols in cloud storage. Both test exact equality, not image similarity.
Research Program
Image preprocessing. Each submitted panel is cut into lanes with the rectangle heuristic of the Learning-to-identify pipeline. Each lane is canonicalised against flips and small rotations, and its intensities are replaced by their ranks. Rank replacement makes the descriptor invariant to any monotone brightness or contrast curve. A band mask removes band pixels. Mask labels are available in the synthetic-blot dataset, which ships binary blot-masks.
Texture encoder. A small CNN reads only the masked background and is trained contrastively. Positives are augmentations of the same scan: offset crops, rescaling, JPEG, gamma changes and partial erasing. Hard negatives are background patches from different membranes probed for the same target.
Tag generation. Each lane embedding is quantised into $L$ locality-sensitive bands. Each band code $b_\ell$ becomes a tag $t_\ell = F_k(\ell \,\|\, b_\ell)$. Here $F_k$ is a pseudorandom function keyed by a key server that evaluates it obliviously, so the key server never sees $b_\ell$, and that rate-limits each institution.
Matching and escalation. Two lanes are flagged when at least $\tau$ of their $L$ tags coincide. A flag triggers human escalation under a disclosure policy, not automatic disclosure.
Leakage analysis. We measure what the server learns by estimating the min-entropy of band codes under an attacker who knows the protein target, molecular weight and imaging platform. That attacker can render candidates with the generative models from the forensic paper. We also measure the residual frequency signal from collision counts.
Baselines
The naive reference point is convergent encryption on file bytes, the key-from-hash scheme used in Feasibility study on security deduplication of medical cloud privacy data. It is run on the same submissions with no tuning. Any recall above zero on re-saved or re-cropped copies counts as beating it, and we expect it to score near zero.
The strongest accuracy competitor is the plaintext pipeline of Learning to identify image manipulations in scientific publications, which reports 90.86% accuracy on its Bik-derived pairs. We retrain it on our training split with the same augmentations and the same hyperparameter search budget as our encoder. We also run the triplet embedding of Image Forensics the same way; no reported result was found on Lacuna for its numeric accuracy on the 126 real cases. These plaintext systems represent the trusted-clearinghouse option. Our private system "wins" if it stays within a pre-registered 3 points of their recall at equal false-positive rate, and if it beats them on hard negatives drawn from different membranes.
The privacy-relevant competitor is the LBPH perceptual-hash scheme of Image Deduplication Based on Hashing and Clustering in Cloud Storage, run with its hash in the clear; no reported result was found on Lacuna for its accuracy.
The key ablation keeps our full keyed-tag protocol but feeds it band-region embeddings instead of background texture. This isolates the texture ingredient on both accuracy and attack resistance. A second ablation removes the rank transform.
Why not just…?
Why not just hash the image files with convergent encryption? Any re-export, metadata change or one-pixel crop changes every byte of the digest. Both the LBPH paper and SimEnc document this failure for images and slightly modified files. We still include it as the floor so that the gap is measured rather than asserted.
Why not just have a trusted clearinghouse run a Siamese network on plaintext scans? This is a real threat to the proposal's relevance. If institutions accept such a party, our system only has to match its accuracy to be pointless. Experiments cannot settle the sociology. What they can do is price privacy precisely: the recall gap between our tags and the retrained plaintext baselines at fixed false-positive rate. If that gap is small, institutions that refuse a clearinghouse lose little by using the private alternative.
Why not just put the existing band-content embedding behind the same keyed hash? This is the closest simpler design and may work. The ablation in the Baselines section decides it. If band-content tags match texture tags on recall, show no excess false positives on loading controls, and yield comparable attacker success under guided guessing, the texture premise is wrong and we would report the simpler scheme.
Why not just compute encrypted distances with homomorphic encryption or pairing-based equality tests? These give stronger guarantees. However, the cited constructions test exact equality, and their pairing and re-encryption costs already bottleneck exact-match systems. We would benchmark one homomorphic distance computation per candidate pair as a confirmation stage on flagged pairs only, not as the index.
Proposed Experiments
We use three datasets. The first is the synthetic-and-real blot dataset of 14,200 real 256×256 crops derived from about 300 originals. Crops sharing an original are natural same-membrane positives, and we split by original to avoid leakage between training and test. Its generated images serve as non-physical negatives. The second is the Bik-derived pairs. The third is the 126 Retraction Watch/PubPeer cases. We also collect raw repeat-experiment scans of common targets from partner labs as hard negatives; no such public set appears in the file.
The metrics are lane-level recall at 1% false-positive rate under journal-style processing (resizing, JPEG, gamma, splicing), plus false-positive rate on same-target, different-membrane pairs. For privacy we report the attacker success rate: the fraction of private tags hit within a budget of oblivious-function queries, using guesses rendered by DDPM and StyleGAN2-ADA conditioned on the true band layout.
Decisive checks.
- Texture recall must survive resizing. The forensic paper flags resizing as the weak point for texture statistics, so we sweep downscaling factors and report where recall collapses.
- Attacker success against texture tags must be orders of magnitude below that against band-content tags at equal recall.
- Generated blots must not collide with real ones. Collisions would indicate that the generator memorised training membranes, a finding we would report either way.
References
- Learning to identify image manipulations in scientific publications
- Image Forensics: Detecting duplication of scientific images with manipulation-invariant image similarity
- Forensic Analysis of Synthetically Generated Western Blot Images
- Image Deduplication Based on Hashing and Clustering in Cloud Storage
- SimEnc: A High-Performance Similarity-Preserving Encryption Approach for Deduplication of Encrypted Docker Images
- Secure Deduplication of Encrypted Data: Refined Model and New Constructions
- Information Leakage in Encrypted Deduplication via Frequency Analysis
- Attribute-Based Equality Test Over Encrypted Data Without Random Oracles
- Private data deduplication protocols in cloud storage
- Feasibility study on security deduplication of medical cloud privacy data