When people post photos online, commercial face recognition systems can match those photos to their identity without consent. The main technical countermeasure is the "adversarial cloak": before posting, a user adds a small, near-invisible perturbation to their photo that pushes the image's embedding — the vector a recognition network computes from a face — away from the user's true identity, so the system fails to match. Because users cannot access the internals of, say, Amazon's recognition service, the perturbation must be computed on a surrogate model the user trains or downloads, and must transfer to models it has never seen. The field has made per-image cloak generation fast (some tools now take seconds, or produce one mask per person reusable across all their photos), but one cost has gone unexamined: preparing the surrogate itself. Today's best tools lean on ensembles of large, fully trained face recognition backbones, each trained on millions of faces. This proposal asks a simple question: does the surrogate need to be fully, standardly trained at all? Recent work on classification benchmarks shows that lightly adversarially trained surrogates — trained with a tiny perturbation budget or stopped early — produce attacks that transfer dramatically better, because standard training bakes in fragile, model-idiosyncratic features. Nobody has tested whether this carries over to face recognition, where the loss function and the geometry of the problem are different. We propose to find out, and to build a privacy cloaking pipeline whose total cost — surrogate preparation plus per-image generation — is a fraction of current tools, with equal or better protection against unknown recognition models.
Three literatures motivate this. On the privacy side, tools like LowKey, which reduced Amazon Rekognition Rank-1 accuracy from 93.7% to 0.6% using an ensemble of standardly trained backbones with differentiable alignment and Gaussian smoothing in the loop, explicitly leave open "whether surrogate models can be made more transferable via lightweight adversarial training, data pruning, or robust tokenization" (LowKey: Leveraging Adversarial Attacks to Protect Social Media Users from Facial Recognition). On the transferability side, DFANet showed that for face recognition, attacking intermediate features rather than final labels is what makes cloaks transfer, and released the TALFW benchmark on which ArcFace's LFW accuracy falls from 99.82% to 63.45% (Towards Transferable Adversarial Attack against Deep Face Recognition). On the training-cost side, the Early Stop paper showed on CIFAR-10 and ImageNet that light adversarial training (perturbation budget $\epsilon = 0.05$) or early stopping purges "Very Non-Robust Features" from a surrogate and lifts black-box transfer — for ResNet18 against a ViT-B/16 target, from 19.4% to 87.4% (Early Stop And Adversarial Training Yield Better surrogate Model: Very Non-Robust Features Harm Adversarial Transferability). But that paper never touches open-set metric learning, face verification, or commercial APIs. The gap sits exactly at the intersection of all three directions, and two other papers flag it from their own sides: Q-TART, which cuts adversarial training cost by pruning unstable samples but only studies defensive resistance to transferred attacks on closed-set classification (Q-TART: Quickly Training for Adversarial Robustness and in-Transferability), and ROB tokens, which make a frozen backbone robust by training only ten prepended token vectors in about 200 steps, but only on ImageNet classification and segmentation (Robustness Tokens: Towards Adversarial Robustness of Transformers).
Core Claim
The claim is that for face privacy cloaks, the surrogate's training recipe — not its size, not the number of ensemble members — is the main lever on both transferability and cost. Specifically: applying light adversarial training, early stopping, and unstable-sample pruning to surrogates trained with open-set metric learning losses (ArcFace, CosFace) will produce cloaks that transfer at least as well to unseen recognition models — including transformer-based ones, where transfer from CNN surrogates is known to degrade — while cutting surrogate preparation compute by a large factor. The mechanism hypothesis, imported from the Very Non-Robust Feature account and from the information-bottleneck view of adversarial training (Adversarial Training Reduces Information and Improves Transferability), is that standard metric-learning training lets each surrogate latch onto high-frequency identity idiosyncrasies; cloaks that exploit them fool that one model but not others. Light adversarial training suppresses those idiosyncrasies, so the remaining cloak directions live in a shared, identity-semantic subspace that other architectures also use. Whether this holds in embedding space — where angular margins, not decision boundaries, define the geometry — is precisely the open empirical question, and we design the study so it can be answered cleanly either way.
Research Program
The program has three parts. First, a diagnostic: train ArcFace and CosFace surrogates on CASIA-WebFace and a VGGFace2 split under four recipes — standard training, early stopping, light adversarial training (a few PGD steps at small $\epsilon$ inside the metric loss), and Q-TART-style pruning of samples whose embeddings move most under Gaussian noise — and measure how each recipe changes the embedding geometry (within-identity dispersion, between-identity angular margins, and the cosine alignment of cloak gradients across independently trained models). Clean verification of the surrogate is monitored throughout; if a recipe drops the surrogate's own LFW accuracy well below the 99.82% reported for standard ArcFace (Towards Transferable Adversarial Attack against Deep Face Recognition), the surrogate is unusable and the recipe is discarded. Second, the cloak pipeline: plug each surrogate recipe into a fixed, strong generation stack — DFANet-style feature-level objectives with convolutional dropout, LowKey-style differentiable detection and alignment, and Gaussian smoothing for compression survival — and measure black-box protection success against a held-out suite of targets the surrogate never saw, plus the commercial APIs used by prior work. Third, a cheap-diversity mechanism: instead of training multiple full backbones for an ensemble, freeze one public backbone and train several small ROB-token sets on top of it, each in a few hundred steps as reported for ROB tokens (Robustness Tokens: Towards Adversarial Robustness of Transformers), treating each tokenized variant as an ensemble member. This borrows a defensive technique for an offensive purpose nobody has explored: parameter-efficient diversity. Finally, we run the arms-race check in both directions, evaluating whether cheaply robustified targets (light-AT face recognition models) resist our cloaks — relevant because BPFA already reports its attacks hold up against adversarially robust face recognition models (Improving Transferability of Adversarial Examples on Face Recognition with Beneficial Perturbation Feature Augmentation).
Proposed Experiments
Evaluation uses the benchmarks established by the closest prior work. Transfer and evasion are measured on LFW and TALFW (Towards Transferable Adversarial Attack against Deep Face Recognition); identity-level, reusable masks are evaluated on Privacy-Commons and Privacy-Celebrities following OPOM (OPOM: Customized Invisible Cloak towards Face Privacy Protection), with OPOM's convex-hull formulation run over both standard and light-AT surrogates; per-image visual quality is audited on CelebA-HQ, where GIFT reports an average Protection Success Rate of 89.73% with FID 31.19 (Transferable Adversarial Facial Images for Privacy Protection). Demographic bias introduced by sample pruning is audited with FairFace-style attribute breakdowns, since Q-TART's own authors flag that pruning unstable samples may disproportionately remove minority identities (Q-TART: Quickly Training for Adversarial Robustness and in-Transferability). For pose and age robustness we use CFP-FP and AgeDB as in APF (Adversarial Privacy-preserving Filter); no reported result found on Lacuna for either benchmark's numeric scores, so we report our own. Metrics are Protection Success Rate against held-out targets, verification true-positive rate reduction, SSIM and LPIPS under a fixed imperceptibility budget, and — the headline — Protection Success Rate per GPU-hour of total system cost (surrogate training plus generation), with per-image wall-clock reported alongside (LowKey's 32 seconds per image on an RTX 2080 Ti is the reference to beat on cost). The decisive comparisons are all compute-matched: every method gets the same data, the same tuning grid, and the same total GPU-hour budget, and is judged on targets excluded from its optimization.
Baselines
The "do nothing" reference point is posting clean photos: with no cloak, ArcFace verifies LFW pairs at 99.82% as reported by the DFANet paper (Towards Transferable Adversarial Attack against Deep Face Recognition), and any serious cloak must drive held-out target verification dramatically below that. The strongest existing system is LowKey (LowKey: Leveraging Adversarial Attacks to Protect Social Media Users from Facial Recognition), which combines a heterogeneous ensemble of standardly trained backbones with pipeline-aware differentiable alignment and Gaussian smoothing, and reports dropping Amazon Rekognition Rank-1 accuracy from 93.7% to 0.6% and Microsoft Azure Face to 0.1%, at 32 seconds per image. It is the strongest competitor because it owns the best published numbers against commercial black-box APIs and its design — heavier standard surrogates — is exactly the alternative our claim contradicts. We will reimplement its pipeline in our harness, run it on the same LFW/TALFW/Privacy-Commons splits with the same LPIPS budget and the same tuning grid over $\epsilon$ and iteration count, and give it the same GPU-hour allowance for surrogate preparation. Beating it means higher Protection Success Rate against held-out academic targets it never saw (especially transformer-based recognizers) at equal or lower total compute, or equal protection at a large compute discount; matching it at one-fifth the cost would already be a positive result. A second competitor is BPFA (Improving Transferability of Adversarial Examples on Face Recognition with Beneficial Perturbation Feature Augmentation), which perturbs the surrogate's internal feature maps during attack generation and reports attack success rates close to 100% on LFW and CelebA-HQ; it is the strongest per-image transferability booster, and we run it on top of both standard and light-AT surrogates so the surrogate recipe is the only varying ingredient. Finally, the key ablation isolates exactly that ingredient: one cloak-generation codebase, one evaluation suite, two surrogates differing only in training recipe (standard versus light adversarial training). If the light-AT surrogate does not raise held-out Protection Success Rate at matched perceptual budget, the core claim is false and we will report it as such.
Why not just…?
Why not just train a bigger standard ensemble? This is the real threat, and it is essentially LowKey's answer. The evidence against it is indirect but substantial: the Early Stop paper found that standard training accumulates model-idiosyncratic fragile features that actively harm transfer, and that light adversarial training lifted ResNet18-to-ViT-B/16 transfer from 19.4% to 87.4% on ImageNet — a cross-paradigm gap that adding more standard CNN members does not close (Early Stop And Adversarial Training Yield Better surrogate Model: Very Non-Robust Features Harm Adversarial Transferability). Still, the classification result might not carry to embedding spaces, so we settle it empirically: the compute-matched comparison in the experiments section gives the ensemble route every advantage (same budget, same tuning), and if the standard ensemble matches light-AT surrogates on both CNN and transformer targets, the proposal fails honestly.
Why not just optimize harder per image? BPFA-style feature augmentation already reports attack success close to 100% on LFW and CelebA-HQ, so a skeptic could argue surrogate engineering is beside the point (Improving Transferability of Adversarial Examples on Face Recognition with Beneficial Perturbation Feature Augmentation). But those numbers come with unexamined costs: iterative per-pair optimization loops, sensitivity to the feature-perturbation step size that the authors themselves flag, and surrogate ensembles that still must be fully trained and maintained. Our claim is explicitly cost-normalized, not "highest ASR at any price." Moreover, BPFA is composable with our ingredient rather than competing with it — we run BPFA atop both surrogate types, and if BPFA-plus-standard matches BPFA-plus-light-AT at equal cost, that ablation refutes us. We consider that a genuine possibility and design it as a decisive check, not a footnote.
Why not just amortize with universal or generative cloaks? OPOM already achieves roughly 86% Top-1 black-box protection on Privacy-Commons with zero-second per-image inference via one reusable identity mask, and GIFT reaches 89.73% Protection Success Rate with photorealistic edits (OPOM: Customized Invisible Cloak towards Face Privacy Protection; Transferable Adversarial Facial Images for Privacy Protection). If a user only cares about per-image latency, these largely solve it — we concede that. But neither touches our target: OPOM's offline convex-hull optimization over the surrogate ensemble remains compute-heavy and its authors flag mask leakage, while GIFT's authors explicitly leave robustness against adversarially trained recognizers to future work and rely on standard surrogate ensembles. Our contribution sits one layer down — the surrogate itself — and is designed to compose with both: we will run OPOM's hull construction over light-AT surrogates as one experiment, predicting cheaper offline optimization and better cross-architecture transfer. If the composed system shows no gain over OPOM with standard surrogates, the surrogate-recipe lever is irrelevant in the identity-mask regime and we will say so.
Decisive Checks and Risks
Three checks decide the proposal. First, the geometry diagnostic must show that light metric-learning adversarial training actually changes cloak-gradient alignment across independently trained models — without this, any downstream gain is luck. Second, the compute-matched curve of Protection Success Rate versus total GPU-hours must show light-AT surrogates on or above the standard-ensemble frontier, with the sharpest test being transfer from CNN surrogates to transformer-based face recognizers, the known weak spot of cross-model transfer. Third, the arms-race check: if cheaply robustified targets neutralize our cloaks, the practical story shifts from attacker-side to defender-side, and the same experiments quantify that trade-off. The largest scientific risk is that angular-margin losses, which already impose large separation between identities, behave unlike classification losses under light adversarial training — possibly the fragile features the VNRF hypothesis targets are already suppressed, or possibly light perturbations collapse identity discriminativity and make the surrogate useless. The monitoring rule above (surrogate LFW accuracy held near standard levels) catches the failure early, and the diagnostic experiment is deliberately scheduled first, so a negative answer costs weeks, not the whole program. If the hypothesis holds even partially, the payoff is a privacy tool whose full preparation cost drops from training-scale to fine-tuning-scale — in the spirit of ROB tokens' 200-step convergence and Q-TART's 17.9% training-time reduction — while protecting users against the recognition systems they cannot see, query, or inspect.
References
- LowKey: Leveraging Adversarial Attacks to Protect Social Media Users from Facial Recognition
- Towards Transferable Adversarial Attack Against Deep Face Recognition
- Early Stop And Adversarial Training Yield Better surrogate Model: Very Non-Robust Features Harm Adversarial Transferability
- Q-TART: Quickly Training for Adversarial Robustness and in-Transferability
- Robustness Tokens: Towards Adversarial Robustness of Transformers
- Adversarial Training Reduces Information and Improves Transferability
- Improving Transferability of Adversarial Examples on Face Recognition with Beneficial Perturbation Feature Augmentation
- OPOM: Customized Invisible Cloak Towards Face Privacy Protection
- Transferable Adversarial Facial Images for Privacy Protection
- Adversarial Privacy-preserving Filter