Cameras that see at night — surveillance systems, driver assistance, doorstep doorbells — face two problems at once. First, deep vision models trained on daylight data fail badly on underexposed images: a face detector that scores over 90% on normal imagery drops to 16.1% mAP on the DARK FACE benchmark of real nighttime photos HLA-Face: Joint High-Low Adaptation for Low Light Face Detection. The standard fix is a small pre-processing network placed in front of a frozen detector, which brightens the image back into the distribution the detector was trained on Zero-Reference Deep Curve Estimation for Low-Light Image Enhancement, Toward Fast, Flexible, and Robust Low-Light Image Enhancement. Second, such systems increasingly learn collaboratively: clients train locally and send only model updates to a server. But gradient inversion attacks show that those updates can be numerically reversed to reconstruct the private training images — an attacker who knows the model and intercepts a gradient can search for a dummy image whose gradient matches the real one Deep Leakage from Gradients, Inverting Gradients - How easy is it to break privacy in federated learning?.
This proposal argues that these two problems are secretly the same dial. The exact property that makes enhancement useful — restoring edge contrast so early convolutional filters fire strongly — is the property that makes gradients information-rich and invertible. Nobody has measured this. The inversion literature evaluates only on well-exposed data (ImageNet, CIFAR, FFHQ), and the low-light literature never considers federated training Gradient Inversion with Generative Image Prior, Self-Guided Low Light Object Detection Framework. We propose to first map the utility–leakage frontier across pre-processing choices, and then push it: train the pre-processor not only to restore detection accuracy but to shape the downstream gradient so that inversion fails, turning a 79,000-parameter illumination module into a client-side privacy mechanism that costs nothing at the server.
Core Claim
The claim has two parts. First, an empirical one: low-light degradation is not free privacy. Raw dark images fed into a detector produce sparse, compressed gradients that defeat naive pixel-space inversion with natural-image priors, but they still leak labels and coarse scene structure, and a modern attack that meta-learns its image prior from intercepted gradients — GIML in Gradient Inversion with Generative Image Prior — should adapt to the dark manifold and recover far more. Meanwhile, standard enhancement pre-processing makes leakage worse in the way that matters: the attacker does not need to reconstruct the noisy dark pixels; they reconstruct the enhanced image, which is a clean, well-exposed, immediately recognizable version of the private scene. Second, a constructive one: because the pre-processor sits on the client and its parameters are never transmitted, it is the one component of the pipeline the adversary cannot directly observe, and it can be trained adversarially against a simulated inverter to buy privacy at a small utility cost — a trade-off noisier defenses like differential privacy pay for much more expensively.
The threat model is the passive "honest-but-curious" server from the systematizations of SoK: On Gradient Leakage in Federated Learning and SoK: Gradient Leakage in Federated Learning: the server follows the protocol, knows the global model weights $g_\phi$, and sees the transmitted gradient $\nabla_\phi \mathcal{L}$, but the client-side enhancer $f_\theta$ (and its update, which we simply never send) is private.
Research Program
Three stages, each building on the last.
Stage 1 — the frontier measurement. We construct federated clients over the nighttime benchmarks in the file: DARK FACE (6,000 real night images), ExDark (7,363 low-light images, 12 classes), and the 4,485-image nighttime subset of nuImages. Each client runs a frozen or locally fine-tuned detector (DSFD-style face detector, Faster R-CNN-style object detector) on one of four input pipelines: raw dark input; classical fixed tone mapping (gamma, histogram equalization); a Zero-DCE-style curve estimator Zero-Reference Deep Curve Estimation for Low-Light Image Enhancement; and a SCI-style self-calibrated block Toward Fast, Flexible, and Robust Low-Light Image Enhancement. The attacker intercepts the transmitted detector gradients and runs four attacks of increasing strength: DLG with Euclidean matching Deep Leakage from Gradients; the cosine-similarity plus Total Variation attack of Geiping et al. Inverting Gradients - How easy is it to break privacy in federated learning?, which we know recovers about 14 dB PSNR from a trained ResNet-20 where Euclidean matching manages only about 7 dB; GIAS with a StyleGAN prior trained on well-exposed faces; and GIML, which meta-learns the prior from scratch Gradient Inversion with Generative Image Prior. The critical comparison is what the prior mismatch does: GIAS's generative prior, trained on well-lit FFHQ or ImageNet, has essentially no support on underexposed, noisy night imagery, since standard datasets contain under 2% low-light images Getting to know low-light images with the Exclusively Dark dataset. We predict it will hallucinate a plausible daytime face rather than the true scene — a failure that is itself dangerous, because it produces confident but false evidence — while GIML should close the gap. We score reconstructions both against the raw dark image (how much of the private signal escaped) and against the enhanced image (how recognizable the leaked content is), plus label recovery in the iDLG sense.
Stage 2 — the inversion-aware enhancer. We train a curve-based enhancer $f_\theta$ with a dual objective. Let $\hat{x} = f_\theta(x_{\text{dark}})$ be the enhanced image and $\nabla_\phi \mathcal{L(\hat{x})}$ the gradient the detector produces on it. We unroll $K$ steps of a differentiable proxy attacker $\mathcal{A}$ (cosine matching with a TV prior, following Geiping) starting from noise, obtaining a reconstruction $\mathcal{A}(\nabla_\phi \mathcal{L})$, and train the enhancer so the reconstruction fails while detection succeeds:
$$ \min_\theta \; \mathcal{L}_{\text{task}}\big(g_\phi(f_\theta(x_{\text{dark}}))\big) \;+\; \lambda \, \text{Sim}\big(x_{\text{dark}},\, \mathcal{A}(\nabla_\phi \mathcal{L}(f_\theta(x_{\text{dark}})))\big), $$
where $\mathcal{L}_{\text{task}}$ is detection loss through the frozen detector plus the non-reference enhancement losses (exposure control, color constancy, illumination smoothness) that let Zero-DCE train without paired data, and the second term rewards similarity between the attacker's reconstruction and the true dark image — the attacker, missing the point, converges to an image that matches the gradient but is uselessly far from the private scene. The intuition is that the enhancer learns to restore the specific feature statistics the detector needs while injecting ambiguity that a natural-image prior resolves wrongly. Because the enhancer is tiny (Zero-DCE is about 79,000 parameters; SCI's inference block is 300), this stays inside the compute budget of an edge camera.
Stage 3 — generalization and combination. We test the trained enhancer against attacks it was not trained against (GIAS with dark-adapted priors, GIML, active attacks of the kind surveyed in SoK: On Gradient Leakage in Federated Learning), across client batch sizes $B \in \{1, 8, 32\}$, since batch averaging is the main source of inversion ambiguity, and we combine it with light differential privacy to test whether the two mechanisms stack — whether an enhancer that already degrades inversion lets us shrink the noise budget needed for a formal guarantee.
Baselines
The do-nothing reference point is the raw pipeline: the detector consumes the unenhanced dark image and the client transmits its gradients. Its utility anchor is the number reported by HLA-Face: Joint High-Low Adaptation for Low Light Face Detection: DSFD collapses to 16.1% mAP on DARK FACE without adaptation, versus over 90% on daylight WIDER FACE. This reference matters for privacy too, because if raw dark gradients already leak under a strong attack, the "darkness as natural privacy" argument is dead on arrival, and Stage 1 measures exactly that.
The strongest attack baseline is GIAS from Gradient Inversion with Generative Image Prior: latent-space inversion through a pretrained generator, which reports around 15.58 dB PSNR where pixel-space attacks achieve 13.78 dB, still reconstructs at batch size 32, and survives 99% gradient pruning. We run it exactly as published, with an FFHQ-prior generator and with a dark-finetuned generator, on the same intercepted updates, iteration budgets, and hardware as every other attack, because the interesting scientific question is what the prior mismatch does rather than whether we can out-tune it. The Geiping cosine-TV attack Inverting Gradients - How easy is it to break privacy in federated learning? is the second attack baseline, since it defined the direction-matching formulation that makes inversion work on trained, deep networks (14 dB on a trained ResNet-20, and reconstruction from a ResNet-152 on 224×224 ImageNet).
The strongest defense competitor is utility-preserving enhancement without any anti-inversion training: a Zero-DCE or SCI pre-processor tuned for detection accuracy alone, run on identical clients with identical compute. Beating it means holding detection mAP within about two points while cutting the best attack's reconstruction PSNR by at least 3 dB on the same intercepted updates. We also run the standard defenses at matched settings — Gaussian noise on transmitted gradients across a $\sigma$ sweep, gradient pruning at 20% (which stops DLG per Deep Leakage from Gradients) and at 99% (which GIAS shrugs off per Gradient Inversion with Generative Image Prior) — with the utility cost of each measured on the same detection benchmarks, since Zhu et al. report that noise magnitudes above roughly $10^{-2}$ are needed to stop reconstruction and that these severely degrade accuracy.
The key ablation isolates the proposal's single new ingredient: the anti-inversion term. The utility-only enhancer and the inversion-aware enhancer share architecture, data, training schedule, and compute; they differ only in the second objective term. Any privacy gap between them is attributable to that term and nothing else. A second ablation removes the enhancer's privacy-by-obscurity assumption by giving the attacker a copy of the enhancer (known $\theta$), which bounds how much of the defense depends on the attacker not knowing the pre-processing rather than on the adversarial shaping.
Why not just…?
Why not just add differential-privacy noise to the gradients? This is the real threat to the constructive half of the proposal, and we treat it as the primary defense baseline rather than arguing it away. The file's evidence says the trade-off is steep: Deep Leakage from Gradients reports that noise on the order of $10^{-2}$ or more is needed to eliminate reconstruction and that this severely degrades downstream accuracy. Our honest position is that the enhancer is not a replacement for DP but a way to move along the curve: if the enhancer already degrades inversion, the formal guarantee should arrive at a smaller $\sigma$ and a higher mAP. Stage 3 tests exactly this by sweeping $\sigma$ with and without the inversion-aware enhancer; if the utility-matched DP curve dominates our method everywhere, the constructive claim is wrong and the measurement claim stands on its own.
Why not just skip enhancement and rely on darkness as natural privacy? Stage 1 is designed to answer this empirically rather than by assertion. The optimistic reading — that underexposure zeroes out early ReLU activations and sparsifies gradients the way 20% gradient pruning stops DLG Deep Leakage from Gradients — has two problems we expect the experiments to expose. GIAS reconstructs images even when 99% of gradient values are removed Gradient Inversion with Generative Image Prior, so sparsity alone is weak evidence of safety; and GIML meta-learns its prior from the intercepted gradients themselves, so it should adapt to a dark manifold that a fixed FFHQ prior cannot represent. There is also a decisive utility counterargument: darkness-as-privacy costs 16.1% versus 44.4% mAP for adapted pipelines on DARK FACE HLA-Face: Joint High-Low Adaptation for Low Light Face Detection. If Stage 1 shows raw dark gradients leak labels and coarse structure anyway, the alternative buys neither privacy nor utility.
Why not just fine-tune the detector on labeled dark data instead of pre-processing? HLA-Face: Joint High-Low Adaptation for Low Light Face Detection reports that DSFD fine-tuned with manually labeled dark-face data reaches only 46.0% mAP on DARK FACE, while self-guided restoration frameworks reach 76.6 mAP Self-Guided Low Light Object Detection Framework — so fine-tuning is both more expensive (nighttime annotation is scarce and costly, with under 2% low-light representation in standard datasets Getting to know low-light images with the Exclusively Dark dataset) and worse. It also does not address privacy at all: Inverting Gradients - How easy is it to break privacy in federated learning? shows depth and training offer no intrinsic protection, recovering 224×224 images from a trained ResNet-152, so a fine-tuned federated detector on raw dark data is simply the do-nothing baseline with better mAP. We include it as a utility ceiling, not a privacy solution.
Why not just use the masked jigsaw patch-scrambling defense? The closest existing defense at this intersection is A Unified Masked Jigsaw Puzzle Framework for Vision and Language Models, which scrambles spatial patches to break the position-embedding gradients that leak spatial layout in transformers. It differs from our proposal in both mechanism and scope: it targets a transformer-specific leakage channel with a synthetic permutation, while we target the continuous, physically grounded transformation of illumination, and we co-design it with the downstream task objective rather than applying a fixed scramble. We will run it as a defense baseline anyway, and we expect it to collapse detection utility, because patch scrambling destroys exactly the local edge structure that nighttime detection depends on and that enhancement exists to restore — the same structure whose restoration, per our core claim, is what makes gradients leak.
Decisive Checks
Three results would settle the proposal. First, the frontier plot: detection mAP on DARK FACE and ExDark against best-attack reconstruction PSNR, across the four input pipelines; if the curve from raw to gamma to Zero-DCE to SCI rises in both coordinates, the "enhancement increases leakage" claim is confirmed, and the widely implicit assumption that dark imagery is private is quantified for the first time. Second, the GIML-on-dark result: whether meta-learned priors close the gap that fixed FFHQ/ImageNet priors leave open decides whether the low-light community can ignore generative attackers. Third, the ablation gap: the inversion-aware enhancer versus its utility-only twin under GIAS and Geiping attacks; holding mAP within two points of the utility-only enhancer while cutting reconstruction PSNR by three or more dB would establish that a 300-to-79,000-parameter illumination module, never transmitted and therefore invisible to the server, is a meaningful privacy control in federated night vision. For transparency: no reported result found on Lacuna for gradient inversion PSNR on any low-light benchmark, on Zero-DCE's absolute DARK FACE average precision, or on SCI's DARK FACE precision-recall numbers — which is precisely the empty quadrant this proposal fills.
References
- HLA-Face: Joint High-Low Adaptation for Low Light Face Detection
- Zero-Reference Deep Curve Estimation for Low-Light Image Enhancement
- Toward Fast, Flexible, and Robust Low-Light Image Enhancement
- Deep Leakage from Gradients
- Inverting Gradients - How easy is it to break privacy in federated learning?
- Gradient Inversion with Generative Image Prior
- Self-Guided Low Light Object Detection Framework
- SoK: On Gradient Leakage in Federated Learning
- SoK: Gradient Leakage in Federated Learning
- Getting to know low-light images with the Exclusively Dark dataset
- A Unified Masked Jigsaw Puzzle Framework for Vision and Language Models