When a company serves a classifier through a paid interface, a competitor can send it many queries, record the answers, and train a copy. This is a model extraction attack, and it recovers 81% to 96% of the victim's performance even when the copy uses a different network design (Knockoff Nets). Two kinds of defense exist. Passive defenses, such as watermarks, only help prove ownership after the theft and are easily removed. Active defenses, such as query monitors or networks that eject suspicious queries early, are designed by hand on top of a fixed backbone. Separately, automated architecture search (NAS, the automatic choice of a network's layers and connections) has become cheap. Zero-cost proxies score a candidate design at initialization with one forward and backward pass. Budgeted estimation finds the shortest training schedule that still ranks designs correctly. Both tools have only ever been aimed at accuracy and latency.
This note asks a simple question that has not been measured: among architectures with equal clean accuracy, does the choice of architecture change how many queries an attacker needs to copy the model? If it does, the cheap-search toolbox can be turned to finding designs that are accurate for paying users but slow to clone, at negligible extra search cost. If it does not, the result is still worth reporting, because it would show that structural defenses against query extraction are a dead end and that effort belongs at the output and monitoring layer.
Core Claim
We hypothesize three things. First, extraction resistance varies substantially across architectures at matched accuracy. Second, this variation can be ranked cheaply, using short training plus small-budget extraction in the style of budgeted performance estimation, without running full attacks on every candidate. Third, the accuracy-maximizing designs that standard zero-cost proxies favor sit at the more extractable end of that spread, because those proxies reward unobstructed information flow.
The deliverable is a measured "extractability" axis for a full NAS benchmark, a cheap estimator for it, and a two-objective search that trades a controlled amount of accuracy for resistance. The search is stress-tested against the hand-designed active defense it would replace or complement.
The closest work is ArchLock. It searches NAS-Bench-201 and TransNAS-Bench-101 with a zero-cost-proxy-trained predictor for architectures that resist being fine-tuned to unauthorized tasks. It explicitly leaves black-box query extraction unaddressed, because an attacker can train an unrelated copy from the outputs alone. We target exactly that threat, which requires a different objective: agreement of an independently trained copy, not transfer accuracy of the stolen weights.
Research Program
The first stage builds ground truth. We take a stratified sample of about 500 architectures from NAS-Bench-201, spanning the accuracy range with dense coverage of the top 10%. The top tier matters because zero-cost proxies are weakest there (EZNAS). We train each architecture on CIFAR-10 with one fixed recipe.
We then attack each trained victim with Knockoff-style extraction (Knockoff Nets) under two query sources:
- Natural out-of-distribution images: downsampled ImageNet.
- In-distribution images: a held-out half of CIFAR-10. This answers the documented weakness of monitors that assume attackers query off-distribution (FDINet; Dynamic Neural Fortresses).
Each attack runs at query budgets of 5k, 20k and 50k, with both soft probabilities and top-1 labels returned. The copy is trained in two configurations:
- Fixed ResNet-18 copy: the attacker uses a standard network of their own choosing.
- Exact-architecture copy: the copy shares the victim's own design. This models the NASPY attacker, who recovers NAS cell topologies through cache side channels and clones to within about 1% accuracy.
We define a victim's extractability as the area under its copy-agreement-versus-query-budget curve. Copy agreement is the fraction of CIFAR-10 test images on which the copy and the victim predict the same label.
The second stage tests whether this property is real or a proxy for something trivial. We regress extractability on clean accuracy, parameter count, average output entropy on the query distribution, and calibration error. We then ask how much variance architecture explains beyond these. The bar is a residual spread larger than the variation across seeds of the same architecture, which we estimate from five repeated runs on 30 architectures.
We also compute the standard zero-cost proxies (Zero-Cost Proxies for Lightweight NAS) and EZNAS-A. With these we test the hypothesis that proxy-favored designs are more extractable at matched accuracy.
The third stage builds the cheap estimator. Following Rethinking Performance Estimation in Neural Architecture Search, we vary victim training epochs, image resolution, extraction query budget and copy training length. We look for the cheapest setting whose extractability ranking preserves Kendall's tau of at least 0.6 against ground truth. In parallel, we fit a small predictor from architecture encoding plus zero-cost proxy values. This predictor is a candidate for a true at-initialization estimate, though we expect it to be weaker, since extractability is a property of the trained function.
The fourth stage runs a two-objective evolutionary search over all 15,625 cells. It maximizes tabulated accuracy and minimizes estimated extractability. We then fully train and fully attack the Pareto-front designs.
Baselines
Naive reference. The do-nothing reference is the highest-accuracy NAS-Bench-201 cell, served unmodified. This is what a provider gets today by using any accuracy-driven search. All extraction gains are reported relative to its copy-agreement curve. Knockoff Nets reports that undefended victims of this kind give up 81% to 96% of their performance; no reported result found on Lacuna for Knockoff extraction specifically on NAS-Bench-201 cells.
Strongest existing active defense. The strongest competitor is Dynamic Neural Fortresses. It adds early-exit classifiers so that suspicious queries leave the network early with misleading answers. It reports surrogate accuracy degraded by up to 12% relative to the MeCo defense, a 2x inference speedup under attack traffic, and 0.47% added parameters. It is the strongest competitor because it already delivers architecture-level extraction resistance without search.
We apply it to the naive reference architecture with its published training procedure. Both it and our searched design receive the same CIFAR-10 data, the same query sources and budgets, and the same copy training. The GPU-hour budget for tuning the defense's exit thresholds matches the GPU-hour budget of our search. Beating it means one of two outcomes at equal clean accuracy (within 0.5 points):
- Standalone win: our searched cell lowers copy agreement more than the naive cell with fortress exits added.
- Stacking win: the searched cell with fortress exits added beats the naive cell with fortress exits added. This shows the effects compound. This is the result we consider most likely to matter in practice.
Output-layer baselines. These alter what the interface returns rather than the network: top-1 label truncation and temperature-scaled probabilities, applied to the naive reference. Knockoff Nets reports that truncation only marginally slows the attack. We nonetheless rerun both, because our in-distribution attacker setting differs from theirs.
Ablation of the key ingredient. The key ingredient is the extractability objective itself. The ablation reruns the identical evolutionary search with extractability replaced by an accuracy-matched random selection, and separately by output entropy alone. If entropy-only search matches the full objective, the "architecture" effect is really a confidence effect, and the claim narrows accordingly.
Why not just…?
Why not just return hard labels or noisy probabilities instead of changing the architecture? Knockoff Nets reports that output truncation only marginally slows extraction. Fortress-style analysis notes that output perturbation still pays the full inference cost of every malicious query. However, temperature scaling is a real threat to our claim. If searched designs resist extraction only because they are less confident, a single post-hoc temperature would reproduce the effect for free. The confidence-matched comparison in stage two and the temperature baseline settle this. We report the claim only for resistance that survives matching output entropy.
Why not just add early exits to whatever backbone accuracy search returns? This is the strongest alternative, and it may be sufficient. The fortress paper itself documents two weaknesses. Its gating fails against in-distribution queries, and its input stem is always executed and therefore exposed. A searched cell attacks a different source of leakage: the shape of the trained function seen by in-distribution queries. Our in-distribution attacker arm and the stacking comparison decide whether the two are complementary. If stacking adds nothing, we will say the fortress approach suffices.
Why not just rank designs with an existing zero-cost proxy, or simply run full extraction on every candidate? Existing proxies were built to predict accuracy. Whether any of them predicts extractability is exactly what stage two measures, and we will use a proxy directly if it reaches the tau bar. Running full attacks on every candidate is how we build ground truth on 500 cells. Doing so for all 15,625, or for larger macro spaces, repeats the cost crisis that budgeted estimation solved by cutting AmoebaNet's 3,150 GPU-days to 0.33 (Rethinking Performance Estimation).
Why not just watermark the architecture to prove theft later? The NAS-based architectural watermark (A Novel Watermarking Framework for Ownership Verification of DNN Architectures) survives fine-tuning but is passive. It is also verified through the same cache side channel that NASPY uses to read cell structure. Most importantly, it does nothing against a copy trained from outputs, which carries none of the victim's structure.
Decisive Checks
The project turns on the first check. If accuracy-matched architectures differ in extractability by no more than seed-to-seed noise, under both copy configurations and both query sources, the structural route is closed. The note then becomes a negative result that redirects effort to monitors and output defenses.
If the spread is real, two further checks decide the search component. The first is whether the cheapest budgeted estimator reaches Kendall's tau of 0.6 against full-attack ground truth, with the top-10% tier reported separately. The second is whether Pareto-front cells, or those cells with fortress exits stacked on top, cut 20k-query copy agreement by more than the fortress defense achieves alone, at clean accuracy within 0.5 points of the naive reference.
For reference on proxy quality, the best accuracy-ranking proxy ensemble reaches Spearman 0.82 on NAS-Bench-201 (Zero-Cost Proxies for Lightweight NAS), and EZNAS-A reaches Kendall's tau 0.65. No reported result found on Lacuna for any proxy's correlation with extractability. Finally, we repeat the main comparison on TransNAS-Bench-101 object classification, where ArchLock found more architectural diversity, to test whether the effect grows with search-space diversity as ArchLock's own results suggest.
References
- Knockoff Nets: Stealing Functionality of Black-Box Models
- ArchLock: Locking DNN Transferability at the Architecture Level with a Zero-Cost Binary Predictor
- EZNAS: Evolving Zero-Cost Proxies For Neural Architecture Scoring
- FDINet: Protecting against DNN Model Extraction via Feature Distortion Index
- Dynamic Neural Fortresses: An Adaptive Shield for Model Extraction Defense
- NASPY: Automated Extraction of Automated Machine Learning Models
- Zero-Cost Proxies for Lightweight NAS
- Rethinking Performance Estimation in Neural Architecture Search
- A Novel Watermarking Framework for Ownership Verification of DNN Architectures