An architecture found by expensive architecture search is itself valuable property. A thief who copies the blueprint (the wiring diagram of layers and operations) and retrains it from scratch on their own data erases every watermark stored in the weights. The only prior defense stamps fixed operations into the blueprint and then checks for them by snooping on a CPU's memory cache while the suspect model runs. That check requires physical access to the suspect's hardware and a full gradient-based search to embed the stamp. We propose a different mark. We use a cheap, training-free search to select an architecture that performs near the top of its search space and also has a built-in, secret sensitivity: when trained on clean data by anyone, it reacts in a predictable way to a family of small input perturbations generated from the owner's private key. Independently designed architectures do not react this way. Ownership is then checked through ordinary prediction queries with a statistical test. No trigger data is ever mixed into training, so the owner pays no accuracy cost from the embedding step itself.
The opportunity comes from an attack paper. The Dark Side of AutoML (EVAS) showed that an evolutionary search can find architectures that learn a trigger's features almost by construction. The search is guided by the conditioning of the neural tangent kernel, a matrix computed at initialization that describes how easily a network can fit a given signal. On CIFAR-10 these architectures reached 81.51% attack success rate, against about 59% for ResNet-18 and about 47–53% for random architectures. They kept 64.57% after being retrained from scratch on clean data. That persistence under cold-start retraining is exactly the property an ownership mark for blueprints needs. EVAS, however, ignored benign accuracy and used the effect only to attack. Our claim is that the same effect can be turned into a defensive, keyed mark. The tax it imposes on task accuracy can then be measured almost for free on tabular benchmarks, which are lookup tables of pre-trained accuracies for every architecture in a space.
Core Claim
We claim three things. First, among architectures whose zero-cost task score places them in the top band of a search space, a zero-cost "key-sensitivity" score varies enough that one can select a member of that band whose clean-trained models respond to the owner's key far more strongly than independently chosen architectures do. Second, this selection costs less than 0.3 points of test accuracy relative to the best architecture the same zero-cost search finds without the mark. Third, the response survives retraining from scratch, fine-tuning and pruning, and can be verified from predicted labels alone. What is new is not architecture watermarking and not the kernel-conditioning signal. What is new is combining a secret key, a utility-constrained zero-cost search, and a hypothesis test against a null distribution of independent architectures, so that the mark is verifiable through a black-box query API.
Relation to Prior Work
Two papers supply the problem framing and the main contrast. A Novel Watermarking Framework for Ownership Verification of DNN Architectures and Ownership Verification of DNN Architectures via Hardware Cache Side Channels fix a random chain of operations in a cell. They then run a full differentiable search (GDAS) around that chain and verify the chain by monitoring cache accesses to BLAS library calls. They report 96.53% clean versus 96.16% watermarked on CIFAR-10, and collision rates below 0.03%. We keep their threat model, in which the blueprint is stolen and retrained. We change both ends of the pipeline. Embedding becomes a training-free search instead of GDAS, and verification uses output labels instead of hardware side channels.
EVAS supplies the core mechanism. Zero-cost Proxy for Adversarial Robustness Evaluation constrains it. That paper builds a training-free robustness score from the kernel's smallest eigenvalue and input-space curvature. It also reports that kernel properties favoring clean accuracy tend to oppose those favoring adversarial resistance. The same tension may appear between accuracy and key sensitivity, and measuring it is part of the proposal rather than an assumption. That paper's "skip-connection bias" warns that initialization-time curvature scores misjudge residual-heavy cells, so we audit our score by skip-connection count.
The utility side reuses established zero-cost scoring. EZNAS evolves a scoring formula with Kendall tau 0.65 on NAS-Bench-201 CIFAR-10. Unifying and Boosting Gradient-Based Training-Free NAS shows that common gradient scores are bounded by the kernel's trace. Its hybrid HNAS reaches 94.07% on NAS-Bench-201 CIFAR-10. Both papers also motivate our choice to constrain search to a top band rather than trust the proxy's ranking at the very top, where the file notes that proxies discriminate poorly.
Research Program
The owner draws a secret seed and commits to it publicly with a hash and timestamp before any search. The seed determines a small perturbation generator: a randomly initialized convolutional network producing bounded, input-dependent perturbations. It also determines a target behavior, namely a pseudo-random target class per input cluster.
The search has two steps and uses no training. First, every candidate architecture is scored with a utility proxy, either EZNAS-A or the kernel trace, and only the top band is kept. Candidates whose proxy-predicted accuracy falls below a tolerance of the proxy-selected best are discarded. Second, within that band, each candidate gets a key-sensitivity score. This is the condition number of the kernel formed by the candidate's initialization gradients on key-perturbed inputs, following EVAS. We compute it relative to the median sensitivity of the same candidate to perturbations from other seeds. The ratio matters: it rewards sensitivity specific to this key rather than generic fragility. Generic fragility would make the mark indistinguishable from an ordinary adversarial weakness and would let a false claimant "find" a key for any model.
Verification works as follows. We query the suspect model on n clean inputs and on their key-perturbed versions, and record the rate at which predictions move toward the key's target classes. We compare this rate against a null distribution. The null is built by training reference architectures from the same space on the suspect's task and measuring their response to the same committed key. The resulting z-test is analogous to the matching-index test in Watermarking GNNs via Explanations. One protocol question must be settled in a pilot before anything else, because the file does not specify it. We do not know whether EVAS's 64.57% cold-start figure used a fixed trigger generator or one re-optimized against the retrained model. Our verification requires a fixed, pre-committed generator, since re-optimizing the key against the suspect would destroy non-repudiation. If a fixed key yields no measurable margin after cold start, the central claim fails. The fallback is a weaker variant that allows a few queries of key adaptation within a small, committed perturbation family.
Baselines
The naive reference point is the unmarked architecture chosen by the same zero-cost pipeline, EZNAS-A or HNAS-style trace scoring, on NAS-Bench-201. Its accuracy defines the utility tax, read directly from the benchmark tables. The strongest reported search result in the file on this benchmark is HNAS at 94.07% CIFAR-10 test accuracy (Unifying and Boosting Gradient-Based Training-Free NAS), against a space optimum of 94.37% (NAS-Bench-201). Beating the tax criterion means the marked architecture's tabular accuracy is within 0.3 points of the unmarked pick on CIFAR-10, CIFAR-100 and ImageNet-16-120, using the same proxy and the same number of scored candidates.
The strongest existing ownership method for blueprints is stamp-based constrained search ( A Novel Watermarking Framework for Ownership Verification of DNN Architectures). It reports a CIFAR-10 drop from 96.53% to 96.16% and collision rates below 0.03% (Ownership Verification of DNN Architectures via Hardware Cache Side Channels Ownership Verification of DNN Architectures via Hardware Cache Side Channels). We reimplement its stamp constraint inside the NAS-Bench-201 cell with GDAS and match the stamp's edge count to our search. We report accuracy, GPU-hours of search, and black-box verifiability. Black-box verifiability is the honest point of comparison. The stamp method should win on false-positive rate when side-channel access exists, and we expect it to have no label-only verification at all. We beat it if we match its accuracy tax at a fraction of its search cost while achieving p < 0.001 from label queries.
For weight-level marks we run WaterMAS on the unmarked architecture with its published settings. On ResNet-18 WaterMAS reports a clean error of 13.58% versus 15.69% unmarked and a roughly 66% accuracy collapse on key removal. It should dominate under fine-tuning and pruning and fail under retraining from scratch, which is the scenario this proposal targets.
The ablation that isolates the key ingredient has three arms. The first uses the utility band plus a random member, with no sensitivity search. The second searches for sensitivity with raw rather than key-specific (ratio) scores. The third is EVAS-style unconstrained sensitivity search, to measure the tax that the utility band prevents.
Why not just…?
Why not just fingerprint the trained model passively, with boundary samples or conferrable adversarial examples? Methods like CITED cost nothing in accuracy, but they fingerprint weights. A blueprint retrained from scratch on new data has new weights and a new decision boundary. We test this directly by applying conferrable-style probes crafted on the owner's model to cold-start retrained copies. We expect them to fall to the reference-architecture level.
Why not just identify the stolen architecture passively, crafting probes that separate the unmarked architecture from references, with no special search? This is the real threat. If an architecture's inductive bias alone produces a distinctive probe response after retraining, the search step adds nothing. The ablation "utility band, random member, key generator optimized post hoc against the owner's own trained model" measures exactly this. If its z-scores after cold start match ours, we will report that the search is unnecessary and reframe the contribution as passive architecture fingerprinting.
Why not just weight-watermark a good architecture with WaterMAS or explanation marks? These resist tampering of the same weights, but retraining from scratch removes them by construction. Explanation marks can additionally be stripped by label-only distillation, per Watermarking Graph Neural Networks via Explanations for Ownership Protection. The comparison under cold start settles this. We also concede the converse limitation: distillation into a different architecture defeats our mark too, so the two approaches are complementary.
Why not just add a watermark loss to a weight-sharing supernet search? Both NAS-Bench-201 and NATS-Bench document that differentiable search collapses toward skip-heavy cells. An auxiliary trigger loss would also reintroduce poisoning into training, which is the source of the performance tax. We include a DARTS plus watermark-loss run at matched GPU budget to confirm.
Proposed Experiments
The main testbed is the NATS-Bench topology space, the same 15,625 cells EVAS searched, on CIFAR-10, CIFAR-100 and ImageNet-16-120. Utility and tax come free from the tables. Security evaluation requires training. For each of 20 keys, we train the selected architecture, its unmarked counterpart, and 50 reference architectures sampled from the top band, each from scratch with 3 seeds on clean data.
The decisive check has three parts. After cold-start retraining, the key's response on the marked architecture must reach p < 0.001 against the reference null within 1,000 label queries. The false-positive rate across held-out references and wrong keys must stay below 1%. The tax must stay under 0.3 points.
We then stress-test the mark under fine-tuning, 90% fine-pruning, and mutating one to two edges of the stolen cell. Mutation matters because in a space this small, near-neighbors of an architecture are plausible independent designs. We will report detection power as a function of edit distance rather than hide this. We also report the correlation between our initialization-time sensitivity score and the trained response, to test whether the zero-cost signal is real, broken down by skip-connection count given the bias reported for curvature proxies. Finally, we transfer the procedure to NDS-DARTS, where the larger space should lower collision risk, and report whether the sensitivity-utility trade-off loosens as expected.
References
- The Dark Side of AutoML: Towards Architectural Backdoor Search
- A Novel Watermarking Framework for Ownership Verification of DNN Architectures
- Ownership Verification of DNN Architectures via Hardware Cache Side Channels
- Zero-cost Proxy for Adversarial Robustness Evaluation
- EZNAS: Evolving Zero-Cost Proxies For Neural Architecture Scoring
- Unifying and Boosting Gradient-Based Training-Free Neural Architecture Search
- Watermarking Graph Neural Networks via Explanations for Ownership Protection
- NAS-Bench-201: Extending the Scope of Reproducible Neural Architecture Search
- A Novel Watermarking Framework for Ownership Verification of DNN Architectures
- Ownership Verification of DNN Architectures via Hardware Cache Side Channels
- WaterMAS: Sharpness-Aware Maximization for Neural Network Watermarking
- CITED: A Decision Boundary-Aware Signature for GNNs Towards Model Extraction Defense
- Watermarking Graph Neural Networks via Explanations for Ownership Protection
- NATS-Bench: Benchmarking NAS Algorithms for Architecture Topology and Size