Spectral Trigger Screens for Architecture Search

A backdoor is a hidden behavior implanted in a model: it classifies normal images correctly, but when a secret pattern — a small patch of pixels, a rare character, a subtle texture — appears in the input, the model outputs whatever label the attacker chose. The usual story is that backdoors are planted by corrupting training data or weights. A newer and more disturbing story is that the vulnerability can live in the architecture itself: The Dark Side of AutoML: Towards Architectural Backdoor Search showed that an adversary can evolve an ordinary-looking cell, built only from standard 1×1 and 3×3 convolutions, that stays backdoored even when a victim retrains it from scratch on perfectly clean data. Meanwhile, automated architecture search evaluates thousands of candidate designs, and nobody can afford to poison-train each candidate to check whether it is a trap.

This proposal argues that the check can be nearly free. A backdoor, at its core, is a spectral phenomenon: to fire on a tiny trigger, the network must amplify a low-energy input pattern into a dominant direction of its internal representations — Spectral Signatures in Backdoor Attacks showed exactly this amplification appears as an outlier projection on the top singular vector of the representation covariance, but only after full training on poisoned data. We propose to look for the seed of that amplification before any training, in how a candidate architecture's feature geometry — its matrix of hidden activations, summarized by singular value decomposition (SVD), which expresses the activations as a set of orthogonal variance directions ordered by strength — responds to a trigger at random initialization. If the response is measurable, it becomes a zero-cost security screen: a milliseconds-per-candidate filter that lets architecture search discard designs that are structurally eager to learn shortcuts, at a tiny fraction of the cost of the only existing alternative, the expensive Neural Tangent Kernel computation used by the attackers themselves.

Core Claim

The claim rests on a chain of two established links that nobody has connected. First, Dextr: Zero-Shot Neural Architecture Search with Singular Value Decomposition and Extrinsic Curvature showed that the condition number of intermediate feature maps at random initialization — the ratio of the largest to smallest singular value, where a low ratio means the activation channels are linearly independent — predicts trainability and accuracy at a cost of 0.07 GPU days for a full search, precisely because it tracks the condition number of the Neural Tangent Kernel (NTK), the matrix that governs how easily a network at initialization can learn an input-output mapping. Second, The Dark Side of AutoML showed that the NTK condition number computed with respect to a trigger generator predicts how susceptible an architecture is to backdoors, reaching an 81.51% attack success rate on CIFAR-10 by minimizing it. If feature-map SVD tracks the NTK (Dextr's link), and the NTK governs trigger learnability (EVAS's link), then a trigger-conditioned feature spectrum at initialization should predict backdoor susceptibility — at forward-pass cost rather than NTK cost. EVAS explicitly leaves the causal mechanism behind vulnerable topologies as an open question and never examines feature eigenspectra; Dextr explicitly evaluates only clean data. Neither has tested the composition, and that is the gap this proposal occupies.

Concretely, for a candidate cell $A$ with random weights $\theta_0$, a small batch of clean inputs $x$, and a trigger template $\Delta$ stamped onto those inputs, we compute a differential spectral score

$$ S(A) \;=\; \frac{1}{L}\sum_{l=1}^{L}\Big[\, d_l(x \oplus \Delta) - d_l(x) \,\Big], $$

where $l$ indexes layers, $d_l(\cdot)$ is a spectral descriptor of the layer-$l$ feature matrix (the log condition number, effective rank, or tail-energy fraction), and $x \oplus \Delta$ denotes triggered inputs. The score asks: how much does this architecture's representation geometry change when a trigger is present, before any learning has occurred? We do not pre-commit to the sign or the best descriptor; both are selected on a held-out split and reported openly, following the multi-descriptor spirit of Z-PEFT: Zero-shot Backdoor Detection in Parameter-Efficient Fine-Tuning via Canonical Spectral Signatures, which discriminates trojaned from clean LoRA adapters from 16 spectral features alone with 0.943 leave-one-attack-out AUROC. Extreme singular values are estimated by power iteration rather than exact SVD, keeping cost per candidate at one forward pass on roughly a dozen images — the same regime in which Dextr searched ImageNet in 0.07 GPU days.

The honest risk is that the chain is suggestive, not proven: random-weight spectra govern inductive bias, as On Random Weights and Unsupervised Feature Learning established, but post-training backdoor behavior could depend on things initialization does not see. The experiments below are designed so that if the signal is absent, we find out cheaply and report it.

Research Program

The program has three parts. The first is building the missing ground truth: no benchmark maps architectures to backdoor susceptibility. We construct a poisoned extension of NATS-Bench — its topology space of 15,625 cells is the natural substrate, since EVAS searched it — by stratified-sampling roughly 2,000 cells across skip-connection densities, training each on CIFAR-10 under a fixed 5% poisoning regime, and recording attack success rate and clean accuracy. Attacks will span three families: a BadNets-style patch trigger, a low-magnitude blended trigger, and an input-aware trigger generator in the style of EVAS. We will reuse the attack and defense machinery of BackdoorBench via its public codebase SCLBD/BackdoorBench so that results are comparable to the broader backdoor literature. The stratification over skip density is deliberate: Zero-cost Proxy for Adversarial Robustness Evaluation documented that initialization proxies systematically mis-rank residual cells, and Architectural Backdoors in Neural Networks shows bypass-style connections are exactly how architectural attacks route trigger pixels past the convolutions. Whether skip density confounds our screen is a first-class experimental question, not a footnote.

The second part is the screen itself: fitting $S(A)$ on half the poisoned benchmark, validating rank correlation with measured attack success rate on the other half, and testing generality with a leave-one-trigger-family-out protocol modeled on Z-PEFT's leave-one-attack-out design. The third part is the payoff: a dual-objective search that maximizes Dextr's clean-accuracy score subject to $S(A)$ falling in its lowest quartile, plus a mechanism study that traces, for EVAS's most and least vulnerable cells, how the triggered representation spectrum evolves from initialization through training. The hypothesis worth testing here, grounded in Chen et al. showing that shortcut learning disperses energy into the small-singular-value tail, is that vulnerable topologies exhibit early tail inflation under triggers at initialization that grows into the post-training amplification that Spectral Signatures detects — which would bridge the "training catch-22" that paper acknowledges and give EVAS its missing causal account.

Proposed Experiments

The correlation experiment asks whether $S(A)$ ranks architectures by measured attack success rate, with Spearman correlation and Kendall's tau reported per trigger family and pooled. The bar comes from the closest precedent: the adversarial-robustness proxy of Zero-cost Proxy for Adversarial Robustness Evaluation achieved a Kendall's tau of 0.33 on its 223-model benchmark — robustness prediction is hard, and matching that bar for backdoors at comparable cost would be a solid result, while the 0.82 Spearman that the voting ensemble of Zero-Cost Proxies for Lightweight NAS reaches for clean accuracy on NAS-Bench-201 is the aspirational ceiling. The cost experiment measures per-candidate wall-clock of our screen against a faithful reimplementation of EVAS's NTK condition number on identical hardware and identical ground-truth labels; the target is equal or better tau at ten times lower cost, motivated by Z-PEFT's demonstration that spectral decomposition of compact objects was roughly 21,000× more memory-efficient than full-weight analysis in its setting. The search experiment reports the clean accuracy of the screened winner against NATS-Bench's ceiling of roughly 94.4% on CIFAR-10 and its measured attack success rate against the random-architecture pool of about 47–53% and ResNet-18's roughly 59% reported by EVAS; success means losing no more than about 1% clean accuracy while landing materially below the random-pool success rate with confidence intervals. Finally, the confound and generality experiments: partial correlation of $S(A)$ with attack success rate controlling for skip density, evaluation within fixed skip-density bins, and a generalization probe on the handcrafted bypass backdoors of Architectural Neural Backdoors from First Principles, whose 100% attack success rate under retraining makes them a severe out-of-distribution test.

Baselines

The naive reference point is selecting an architecture at random from the search space with no screening at all. On the EVAS threat model this is not safe by any means: random NATS-Bench cells reach roughly 47–53% attack success rate on CIFAR-10 under input-aware triggers, and ResNet-18, a sensible default choice, sits at about 59% (The Dark Side of AutoML). The strongest existing competitor is EVAS's own NTK condition-number proxy, which is the only initialization-time statistic in the literature demonstrated to predict backdoor susceptibility — its authors report architectures found by minimizing it reach 81.51% attack success rate on CIFAR-10. We will reimplement it, run it on exactly the same candidates and the same poisoned ground-truth labels, give it the same trigger-generator tuning budget our screen receives, and compare rank correlation and per-candidate wall-clock on shared hardware; beating it means matching or exceeding its tau at an order of magnitude lower cost, or beating its tau outright at equal cost. A second family of baselines are the clean-accuracy zero-cost proxies: Dextr itself, and the Synflow/SNIP/jacob_cov voting ensemble of Zero-Cost Proxies for Lightweight NAS, run to test whether simply seeking accurate architectures already avoids backdoor-prone topologies — our expectation is it does not, but if accuracy-seeking selections turn out to be safe, the security screen is redundant. The final baseline is the ablation that isolates our key ingredient: the identical spectral pipeline with a generic noise perturbation in place of a trigger, and plain Dextr condition numbers with no perturbation at all. If a generic noise differential predicts attack success rate as well as the trigger differential, the trigger-specific ingredient is doing no work, and we would report that and reframe the screen as a general input-sensitivity measure.

Why not just…?

Why not just use pruning-saliency proxies like Synflow or SNIP? They are cheaper still — a single backward pass with no singular values at all. But the evidence is against them carrying security signal: on Tiny-RobustBench, GradNorm, SNIP, and SynFlow all achieved lower Kendall's tau with true robust accuracy than the spectral-plus-curvature proxy, which reached 0.33 (Zero-cost Proxy for Adversarial Robustness Evaluation), and Dextr's SVD condition numbers beat prior proxies by 8% Spearman correlation on NAS-Bench-101 (Dextr). We will nonetheless run them on our benchmark; if they match our tau on attack success rate, the spectral machinery is unjustified and the proposal should die.

Why not just compute the NTK directly, as EVAS already does? This is the most serious alternative, and at NATS-Bench cell scale it is genuinely affordable. Our answer is twofold. Empirically, the full empirical NTK requires gradients of every output dimension with respect to all parameters, and both Dextr (a full ImageNet search in 0.07 GPU days versus DARTS's 4.0 GPU days for a worse result) and Z-PEFT (21,000× memory reduction from spectral decomposition of compact objects) show spectral surrogates deliver comparable signal at a fraction of the cost — which matters exactly when the point is screening thousands of candidates. Scientifically, EVAS treats the NTK condition number as a black-box correlation and names the causal mechanism an open problem; our feature-spectrum measurements are the first attempt to explain why certain topologies are trigger-hungry, which the NTK number alone cannot do. The head-to-head cost-and-tau experiment settles the empirical part either way.

Why not just clean the data, retrain from scratch, or fine-prune? Because the threat model is architectural, these fixes provably fail: EVAS architectures retain 64.57% attack success rate after complete retraining on clean data, where weight-based backdoors drop to near zero, and they hold their success rate even under 90% neuron pruning in the Fine-Pruning protocol; handcrafted architectural backdoors survive retraining at 100% success rate (Architectural Neural Backdoors from First Principles). A defense that inspects data or trained weights cannot see a vulnerability that lives in the wiring diagram, which is precisely why a screen that reads the wiring at initialization is needed.

Why not just train each candidate for a few epochs on poisoned data and watch for emerging backdoor behavior? Budgeted low-fidelity evaluation is the standard NAS fallback, but it hits the documented "GPU gap": proxies using 1/64th the FLOPs of reduced training yield only about a 4× wall-clock speedup due to memory and kernel overheads (Zero-Cost Proxies for Lightweight NAS), and partial-training robust search cost roughly 1 GPU day against 0.017 GPU days for zero-cost scoring (Zero-cost Proxy for Adversarial Robustness Evaluation). Worse, BackdoorBench shows early-epoch backdoor dynamics differ sharply across topologies at identical poisoning ratios, so a short poisoned run may mislead. We will still include a 5-epoch poisoned baseline; if it predicts attack success rate as well as our screen at acceptable cost, we will report the cost-accuracy Pareto front honestly and let the reader choose the operating point.

Why not just inspect the architecture's code? EVAS cells contain no malicious code to find — they are legal sequences of ordinary convolutions whose vulnerability is purely topological, and the user study of Architectural Neural Backdoors from First Principles found ML developers caught hand-coded backdoors in only 37% of inspections and actively preferred backdoored models in 33% of pairwise comparisons. Static review is a complement, not a substitute.

Closing Decisive Checks

The proposal is falsifiable at four checkpoints. The screen lives or dies on the held-out rank correlation between $S(A)$ and measured attack success rate, with Kendall's tau of at least 0.33 — the bar set by the adversarial-robustness proxy — as the pass mark, and the leave-one-trigger-family-out test as the generality mark. The cost claim lives or dies on the head-to-head against the reimplemented EVAS NTK proxy. The practical claim lives or dies on the search experiment: a screened winner within about 1% of the clean-accuracy ceiling whose measured attack success rate falls materially below the 47–53% random-architecture pool. And the science claim — that architectural backdoors are visible as trigger-evoked spectral distortion that grows into the post-training amplification Spectral Signatures detects — lives or dies on the initialization-to-training spectral trace across EVAS's extreme architectures. If the first checkpoint fails, the follow-on questions are moot; if it passes, we will have turned a milliseconds-forward-pass into a security instrument for a threat that no amount of data cleaning or retraining can remove.