Suppose a hospital or a wildlife survey sends images to a cloud service for analysis without wanting the server — or anyone on the network — to see the pictures. The standard tool is homomorphic encryption: the client encrypts each image, the server runs its analysis program directly on the encrypted pixels, and only the client can decrypt the result. This scheme has a well-known structural weakness inherited from encryption itself: an eavesdropper learns nothing about the content, yet can still modify it. Adding a second encrypted texture on top of the encrypted image, slot by slot, produces a valid ciphertext of an image with that texture pasted on — no decryption key needed. Cryptographers call this malleability, and the standard confidentiality definition (indistinguishability under chosen-plaintext attack) says nothing against it. The failure is not hypothetical: Telegram's MTProto was shown to break chosen-ciphertext security precisely because manipulated ciphertexts still decrypt successfully in On the CCA (in)security of MTProto.
The core observation of this proposal is that the best cover for such a tampering attack is visual camouflage. An attacker who injects random noise betrays itself the moment anyone audits the decrypted image; an attacker who injects a natural, background-matching texture does not. Adversarial Camouflage: Hiding Physical-World Attacks With Natural Styles showed that style-transfer-based natural perturbations are rated "natural and realistic" by 80.7% of human raters versus 19% for standard adversarial patches. Merged with ciphertext malleability, camouflage becomes doubly invisible: cryptographically (the attacker never sees the plaintext) and perceptually (any decrypted audit looks like an ordinary image). We study this attack on the hardest segmentation setting there is — camouflaged object detection — and build a defense based on restricting what homomorphic edits the system accepts, following the "targeted malleability" principle that malleability should be a licensed feature, never an unlicensed one.
Core Claim
The claim has two halves, one about the attack and one about the defense. On the attack side: an adversary who has never seen the plaintext (only the public key and the deployment-domain image distribution) can craft a ciphertext modification $c' = c + \mathrm{Enc}_{pk}(t)$ with a domain-matched camouflage texture $t$, such that the segmentation the client decrypts is predictably degraded — the camouflaged target is suppressed — while the tampered image itself survives human audit. The relation is distributional rather than pointwise: the attacker cannot tailor $t$ to the specific image, so "predictable" means the mask transformation holds with high probability over the image distribution, which is exactly the threat malleability definitions care about. On the defense side: a wrapper that authenticates the computation program, not merely the ciphertext, restores integrity — tampering either fails verification or produces an output bearing no useful relation to the original — at modest overhead. The design tension that makes this non-trivial: injecting a texture is a linear homomorphic operation, and so is the first layer of any segmentation network, so authenticating "linear operations are allowed" is provably insufficient; the wrapper must authenticate the exact circuit and value bounds.
Nothing in the literature covers this intersection. The physical-camouflage line — Adversarial Camouflage, Physical Adversarial Camouflage Through Gradient Calibration and Regularization, and related attention-suppression attacks — assumes the attacker sees the image and attacks a classifier or detector in plaintext, with no cryptographic model. Camouflage generation work like To Blend In, First Decouple: Rethinking Camouflage Image Generation via Context-Decoupled Representations optimizes blending quality, not attack surface. Conversely, the cryptography side — Targeted Malleability: Homomorphic Encryption for Restricted Computations, Improved, Black-Box, Non-Malleable Encryption from Semantic Security, and non-malleable codes such as Non-Malleable Secret Sharing against Affine Tampering — treats tampering as abstract algebra on bit strings, never asking whether the tampering direction is perceptually deniable or how a vision decoder responds. This proposal runs both directions through each other on real benchmarks.
Research Program
Three components. First, a formalization: we define a perceptual malleability experiment for encrypted inference in which the adversary, given only the public key and oracle access to the encryption of textures of its choosing, wins if (a) the decrypted segmentation mask is predictably related to the honest mask by a target-suppression relation, and (b) the decrypted image passes a perceptual indistinguishability test against untampered images from the domain. The (b) clause is what visual camouflage buys and what ordinary malleability attacks lack; it is modeled on the human study protocol of Adversarial Camouflage. Second, an empirical attack study. Because additive homomorphic operations are linear, they can be simulated exactly in the plaintext domain — adding $\mathrm{Enc}(t)$ to $c$ is mathematically identical to adding $t$ to the image — so we can study attack power at scale before paying the cost of a full encrypted implementation, then confirm on a real additively homomorphic layer for a subset. We attack four published camouflaged-object-detection architectures spanning the design space: SINet with its texture-enhanced local search Concealed Object Detection, the region/edge dual-stream MGL Mutual Graph Learning for Camouflaged Object Detection, the separable foreground/background attention of CamoFormer CamoFormer: Masked Separable Attention for Camouflaged Object Detection, and the one-step diffusion model CODiff, whose Global Semantic Guidance rests on a single DINOv2 class token CODiff: One-Step Diffusion Model for Camouflaged Object Detection. Our hypothesis, drawn from how these models differ, is that models leaning on local texture cues are far more malleable to localized texture injection than models leaning on global semantics — and that one-step diffusion's compressed conditioning makes it brittle to structured shifts, a question no COD paper has asked. We also measure an attacker-knowledge gradient from fully blind (domain-level texture only) to oracle (texture matched to the specific background), to quantify how much the IND-CPA blindness constraint actually costs the attacker.
Third, the defense: a targeted-malleability wrapper. The client encodes the image across packed slots with per-block authentication derived from secret-sharing-style redundancy, so partial tampering invalidates rather than corrupts — the same "decode from honest partial views, reject tampered components" principle that the single-codeword Reconstructable Probabilistic Encoding construction introduced for non-malleable encryption in Improved, Black-Box, Non-Malleable Encryption from Semantic Security, where we also take its overhead lesson seriously and aim for constant-factor overhead rather than the $\Theta(kn)$ ciphertext blowup of earlier matrix constructions. The verification tag is evaluated homomorphically and remains valid only under the exact whitelisted circuit of the deployed segmentation network with bounded values; anything else — including any added texture ciphertext — fails. This mirrors, in the encryption setting, the simulation-extractability template from proof systems, where a tampered proof either fails verification or yields nothing useful to the tamperer Fiat-Shamir Bulletproofs are Non-Malleable (in the Random Oracle Model): the system's answer to manipulation should be rejection, not a related output.
Proposed Experiments
Datasets and splits follow the published protocol exactly: COD10K with its 2,026 test images, CAMO with its 250 test images, NC4K with its 4,121 images, the 76-image CHAMELEON set, and OPC16K for the negative-sample question. Metrics on the attack side are mask IoU drop and MAE increase on camouflaged targets at fixed injection budget (fraction of slots covered), attack success rate (target erased below a threshold), and an Amazon Mechanical Turk naturalness audit replicating the 130-participant protocol of Adversarial Camouflage, where the bar to beat is that paper's reported 80.7% naturalness rating. For clean segmentation context we anchor to published numbers: CamoFormer reports a weighted F-measure of 0.786 on COD10K (against a previous best of 0.731 for FDNet) CamoFormer, CSGNet reports an S-measure of 0.874 and mean error 0.020 on COD10K Cross-Layer Semantic Guidance Network for Camouflaged Object Detection, and CODiff reports over 22% MAE improvement on COD10K at 0.07 s/image and 254.41 GFlops versus CamoD's 0.18 s and 506.65 GFlops CODiff. For the uncertainty axis we note that UGTR's exact MAE on COD10K has no reported result found on Lacuna, only its qualitative improvement over SINet Uncertainty-Guided Transformer Reasoning for Camouflaged Object Detection, so we will re-measure under our protocol rather than cite a number.
The defense experiments report tampering detection rate, false-reject rate on legitimate whitelisted computation (must stay below 1%), and overhead relative to the plaintext pipeline, with CODiff's 0.07 s/image figure as the latency scale a practical wrapper cannot ignore. The OPC16K experiments ask whether texture injection also induces hallucination on images that contain nothing: Is There Really a Camouflaged Object? Towards Realistic Camouflaged Object Detection showed closed-world models reach false positive rates up to 97.4% on pure backgrounds, and its OPCNet reaches a background false positive rate of 0.050 with three-way accuracy 0.905 and camouflaged-sample MAE of 0.038 — we measure how injected camouflage textures shift these numbers, since a hallucination-prone model gives the attacker a second, cheaper victory: make the client see objects that were never there.
Baselines
The "do nothing" reference is the encrypted pipeline with no attacker, which sets the utility ceiling, plus the naive attack of injecting uniform noise at the same budget — the reference every defense trivially beats on perceptual grounds and no defense should be credited for. The strongest competitor on the attack question is the AdvCam-style natural-style camouflage attack itself Adversarial Camouflage, which is the strongest existing method because it is the only published attack achieving human-level stealth (its reported 80.7% naturalness rating versus 19% for adversarial patches) with demonstrated physical robustness; we run its style-transfer pipeline transplanted into our encrypted-domain setting, at the same injection budgets, with the same AMT audit protocol, and it additionally serves as the oracle-knowledge upper bound for our blind attacker. Beating it means matching its perceptual stealth while succeeding under the stricter knowledge constraint of never seeing the plaintext. On the model side, the strongest victims are CamoFormer and CSGNet, the best-performing published detectors in the file (wF 0.786 and S-measure 0.874 on COD10K respectively), run from public checkpoints on the same test splits, so that malleability rankings are not confounded with base accuracy. The training-time competitor is camouflage-augmented adversarial fine-tuning of each victim, given the same GPU budget as our defense wrapper requires, evaluated at matched injection budgets. Finally, the ablation that isolates the proposal's key ingredient is authentication with a permissive whitelist (any bounded linear operation) versus the exact-circuit whitelist: our claim predicts the permissive version fails — texture injection is itself linear — and that predicted failure is the sharpest evidence that program-level authentication, not generic integrity checking, is doing the work.
Why not just…?
Why not just use standard authenticated encryption (encrypt-then-MAC)? A MAC over the ciphertext detects every modification, including ours, with probability one — but it also rejects every legitimate homomorphic computation, which eliminates the reason the pipeline is encrypted in the first place. This is exactly the trade-off articulated by Targeted Malleability: Homomorphic Encryption for Restricted Computations: the goal is not to remove malleability but to restrict it to designated computations, and our wrapper is that principle instantiated and measured on a real vision workload.
Why not just run the inference inside a Trusted Execution Environment? This is the most serious alternative: Achieving trustworthy Homomorphic Encryption by combining it with a Trusted Execution Environment already combines hardware integrity guarantees with homomorphic encryption, and we treat it as an honest competitor rather than a strawman. It shifts trust to hardware vendors and inherits side-channel exposure, while our wrapper is purely cryptographic and composable with untrusted servers; the decisive experiment is a head-to-head on detection rate, false-reject rate, and overhead. If the TEE route matches our guarantees at lower cost, the proposal's durable contribution narrows to the attack study and the architecture malleability ranking — and we will say so in the results.
Why not just adversarially train the segmentation model on camouflage textures? We run this baseline at matched compute. It is a real threat to the defense's practical value, and Adversarial Camouflage gives reason for caution in both directions: natural-style attacks proved robust across viewpoints and conditions, but more fundamentally, robustness only tolerates tampering — it never detects it, never flags the audit log, and never helps the client know the ciphertext was altered. The experiments quantify the recovery each approach buys and at what budget.
Why not just use PGD or adversarial patches? Two concrete reasons from the literature. Perceptually, standard patches were rated natural only 19% of the time in the AdvCam human study, so a tampering attack that might be audited cannot use them. Computationally, iterative gradient attacks require the attacker to evaluate the model on the plaintext image — access the IND-CPA guarantee explicitly denies; a blind, one-shot, gradient-free, distribution-level camouflage injection is essentially the strongest attack the threat model permits, which is precisely why it deserves systematic study.
Decisive Checks
Two outcomes would falsify the core claims, and we design for them up front. If the fully blind camouflage attack cannot degrade masks by more than uniform noise at equal perceptual stealth, the perceptual-cover half of the claim fails and the work reduces to a known-malleability cautionary note. If the permissive-whitelist ablation defends as well as the exact-circuit whitelist — contradicting the linearity argument — then program-level authentication is not the key ingredient and the defense story must be rebuilt. Conversely, a blind attack matching AdvCam's 80.7% naturalness while erasing camouflaged targets across COD10K, CAMO, and NC4K, together with a wrapper that drives its success rate to near zero under constant-factor overhead, would establish that the two fields have been ignoring a shared vulnerability — and a shared fix — that neither could see alone.
References
- On the CCA (in)Security of MTProto
- Adversarial Camouflage: Hiding Physical-World Attacks With Natural Styles
- Physical Adversarial Camouflage Through Gradient Calibration and Regularization
- To Blend In, First Decouple: Rethinking Camouflage Image Generation via Context-Decoupled Representations
- Targeted malleability: homomorphic encryption for restricted computations
- Improved, black-box, non-malleable encryption from semantic security
- Non-Malleable Secret Sharing against Affine Tampering
- Concealed Object Detection
- Mutual Graph Learning for Camouflaged Object Detection
- CamoFormer: Masked Separable Attention for Camouflaged Object Detection
- CODiff: One-Step Diffusion Model for Camouflaged Object Detection
- Fiat-Shamir Bulletproofs are Non-malleable (in the Random Oracle Model)
- Cross-Layer Semantic Guidance Network for Camouflaged Object Detection
- Uncertainty-Guided Transformer Reasoning for Camouflaged Object Detection
- Is There Really a Camouflaged Object? Towards Realistic Camouflaged Object Detection
- Achieving trustworthy Homomorphic Encryption by combining it with a Trusted Execution Environment