acceptodds
Under review as a conference paper at ICLR 2027

DIABLO & PAWS: Parameter-Space Detection and Obfuscation of Encrypted Backdoors

Abstract

Backdoor attacks cause neural networks to exhibit malicious behaviour when triggered, making their reliable detection a critical supply-chain security priority. Encrypted backdoors are particularly concerning because the malicious behaviour is compiled directly into the model architecture and is provably unelicitable, defeating standard white-box auditing techniques. To address this threat, we introduce DIABLO, a compound anomaly detector that identifies encrypted backdoors via statistical signatures in neural network weight histograms. We then red-team this detection pipeline and propose PAWS, a parameter randomisation scheme that preserves model functionality while obfuscating these signatures, exposing fundamental limitations of anomaly-based encrypted backdoor detection. Finally, we show that architectural choices - particularly tree-based XOR implementations - substantially affect robustness to noise, highlighting the need for defence strategies that account for adaptive, obfuscated backdoors and reinforcing the importance of strong model provenance guarantees.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
encrypted backdoors, AI Security, cryptography, compiled neural networks, hidden capabilities, undetectable threats

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_kuFnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies the detectability and robustness of encrypted neural backdoors. The authors propose DIABLO, a statistical detector based on parameter distributions, and PAWS, a randomized parameter assignment scheme to evade detection while preserving Boolean functionality. They further investigate tree-based XOR implementations to improve robustness against parameter perturbations.

Critical strengths

1.The paper addresses an interesting security problem. Studying both encrypted backdoor detection and adaptive evasion provides useful insights into the limitations of parameter-based detection. 2.PAWS provides an interesting approach to preserving Boolean functionality while randomizing parameter values. The paper is generally well organized and easy to follow.

Critical weaknesses

Limited practical relevance and insufficient experimental validation. The experiments compare standalone compiled backdoor MLPs with benign Transformer models. No backdoor is actually integrated into a pretrained LLM. Consequently, the reported detection and evasion results do not establish their effectiveness on real backdoored LLMs. The robustness evaluation is also limited to reduced Keccak configurations, a single trigger/payload pair, and Gaussian noise, without evaluating the impact on normal model performance.

Weak novelty in detection and insufficient support for key claims. DIABLO relies on simple parameter uniqueness and KS-statistic thresholds. PAWS largely evades this detector by increasing parameter diversity, which directly defeats the detector's predefined uniqueness criterion. This result is somewhat expected, and no stronger detection baselines are evaluated. Moreover, Theorem 3.5 claims distribution alignment without sufficient theoretical justification. PAWS uses sign-constrained sampling and gate-dependent biases, which do not generally guarantee that the resulting parameter mean and standard deviation approximate those of the target model.

Questions for authors

1.Can the authors demonstrate DIABLO and PAWS on an actual backdoored LLM rather than isolated MLP modules? 2.How does PAWS perform against alternative statistical or learning-based anomaly detectors? How does DIABLO handle quantized or sparse benign models? 3.Can the authors provide a rigorous justification for the distribution alignment claimed in Theorem 3.5?

Reviewer_X2HeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper investigates the detection and robustness of encrypted neural backdoors. The authors propose DIABLO, a statistical detector that identifies backdoors using layer-wise parameter distributions and uniqueness. They then introduce PAWS, which randomizes Boolean circuit parameters to evade detection while preserving functionality. Experiments show that DIABLO detects all 26 default backdoors but fails to identify the 25 PAWS-obfuscated backdoors.

Critical strengths

S1. The research direction is meaningful, and investigating whether backdoors that are difficult to trigger through input testing leave detectable traces in model parameters is an interesting problem.

S2. The method is intuitive, and the overall research approach is clearly presented. DIABLO examines parameter value diversity and the distribution of layer-wise parameter means, making its detection mechanism easy to understand. The connections between the subsequent parts of the paper are also clear.

Critical weaknesses

W1. The paper lacks validation in a realistic setting. The experiments are conducted on standalone compiled backdoor MLPs, without showing whether these backdoors can be integrated into pretrained Transformers while preserving normal task performance. The parameter distributions of standalone modules can differ substantially from those of complete models, so it remains unclear whether DIABLO can still detect the backdoors once they are integrated into a normal model.

W2. The evasion results are mainly demonstrated against the authors' own relatively simple detector, which limits the generality of the findings. DIABLO flags a model as high-risk only when multiple conditions are satisfied. PAWS can bypass detection simply by increasing parameter diversity, causing one of these conditions to fail. Therefore, the fact that all 25 obfuscated backdoors evade detection mainly highlights a limitation of DIABLO's current decision rule.

W3. Some of the theoretical claims are stronger than what is currently supported. Theorem 3.5 states that PAWS makes the mean and standard deviation of weights and biases closely approximate those of the target distribution. However, the proof in the appendix mainly relies on the fact that parameters are sampled from the target distribution. In practice, the sampling is subject to sign and magnitude constraints, disconnected weights remain zero, and most gate biases are computed from the sampled weights. These factors can affect the resulting parameter distribution. The paper does not clearly define what "closely approximate" means or specify the conditions and error bounds under which this property holds. Therefore, the current proof does not fully support the claimed distribution alignment guarantee.

Questions for authors

Q1. What practical detection conclusions can be drawn from experiments on standalone backdoor modules? Section 4.1 acknowledges that full integration into pretrained Transformers while preserving normal model functionality remains unresolved. If layer-wise statistics are computed over the entire model, could the large number of benign layers dilute the backdoor signal? What evidence currently supports extending the findings from standalone modules to complete models?

Q2. Can PAWS evade detectors that exploit the remaining statistical differences? Since breaking the parameter uniqueness condition is sufficient to bypass DIABLO, what would the detection rate of PAWS be under a statistical detector that does not rely on this condition, using the same trusted calibration data and false positive constraints?

Q3. How do the thresholds used in the logical correctness proofs relate to those in the robustness analysis? The randomized Boolean gates in the main text use a threshold of 0, while Appendix G.4 uses 0.5 to analyze error tolerance. Is there a consistent bias transformation or implementation convention connecting these two definitions? Which definition is used to obtain the numerical results in Table 1?

Reviewer_bx56OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes DIABLO, a simple detector that detects encrypted backdoors through their unusual weight statistics, and PAWS, a way to hide them by sampling weights from real models. It also shows that a tree-based XOR design makes these backdoors somewhat more robust to weight noise.

Critical strengths
  • The paper looks at the problem from both sides. They show a detector that catches the standard encrypted backdoors, and then they show a cheap way to hide those backdoors.
Critical weaknesses
  • The detector is never tested on an injected backdoor. All results compare standalone backdoor MLPs against layers of real LLMs. The paper admits injection is still an open problem. DIABLO's rule is computed over the whole model's MLP layers. So I dont know if it would still flag a backdoor once it is part of a real model. Also a standalone MLP with step activations already looks different from LLM layers by architecture. This likely makes detection look easier than it would be in practice.
  • Beating DIABLO doesnt prove much. The paper itself says the uniqueness check alone is enough for evasion. That check can be broken just by multiplying each neuron's weights and bias by a random positive number, which keeps the function the same. PAWS also leaves unused weights at exactly zero and computes biases from the weights. The paper doesnt test whether simple sparsity or structure checks would catch this.
  • It is unclear if robustness and obfuscation hold at the same time. The paper doesnt say which weights the noise experiment uses. The "dual benefit" claim needs a direct robustness test on PAWSed backdoors.
  • The robustness tests only use Gaussian noise. Fine-tuning is not tested, even though it is the most likely defence for models shared on Hugging Face.
  • The thresholds are tuned on only a few default backdoors from the same compiler used at test time. It is unclear if DIABLO would generalise to backdoors built in any other way.
Questions for authors
  • Which weights does the noise experiment in Section 5 use for the tree XOR models, unit weights or PAWS weights?
  • Do PAWSed layers keep the same share of exact zeros as default backdoors? How does this compare to the benign models in your test set?
  • Test robustness against fine-tuning.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.