- Rating
- 3
- Confidence
- 2
- Informativeness
- 3
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper introduces a new adversarial objective for cooperative MARL: an attribution-evasive attack. A compromised agent should not only cause team failure but also avoid becoming the most attributable agent in post-hoc attribution analysis. The authors observe that existing attacks optimize only disruption, whereas auditors localize the responsible agent by counterfactual attribution, so a 'high-disruption + high-attribution' attack leaves a diagnostic footprint. Therefore, this paper proposes the AE-APL method, which augments the adversarial PPO objective with an exposure-aware penalty. A centralized Joint-Q critic constructs feasible counterfactual references for each agent, computes the local harmful contribution, normalizes it into attribution shares, and builds share-excess and pairwise soft-rank exposure terms. An attack-capability gate ensures that the penalty is applied only when the attack is strong enough. Experiments in MaMuJoCo tasks and three SMAC scenarios show that AE-APL improves attribution stealth over strong attack baselines while maintaining competitive attack effectiveness.
Critical strengths
(1) The new problem formulation captures the key comparison constraint. Treating attribution exposure as an adversarial objective and insisting on comparison at matched attack strength, avoiding the trivial solution of 'attack less to be less exposed'. This makes the paper more than another attack and it exposes a systematic blind spot in attribution/forensics pipelines and is directly useful for designing more robust auditors. (2) The independence and multi-view nature of the evaluation in this paper is seldom seen in similar works. Seven post-training auditors span counterfactual-value, Shapley/coalition, gradient families, auditor training data which is disjoint from attack training/evaluation, the benign policy which is unavailable during attack training, and simulator replay provides environment-level counterfactual validation. This substantially reduces the possibility of overfitting the training critic.
Critical weaknesses
(1) There is a conceptual gap between 'reduced attribution exposure' and 'reduced true responsibility', and the metrics can be strategically exploited. Normalized rank/Top-1 only measures the compromised agent's position in an ordering. An attacker can 'spread the blame' or exploit the locality of one-step counterfactual attribution to stay inconspicuous under current auditors while remaining the true cause of coordinated failure. Appendix Table 12 shows that attacker raw harm drops from 3.144 to 1.063, but the benign raw-harm sum also drops from 20.711 to 18.608 and the maximum benign raw harm from 9.744 to 8.978, suggesting that the failure is not explained by 'some benign agent becoming more harmful', but that the attacker may have learned to distribute harm through coordination dynamics that existing attribution functions fail to capture. The authors should state explicitly that this is 'evading current attribution methods' rather than 'being less responsible'. (2) The comparison with the strongest baselines is not Pareto-based, and an external control optimized for the same objective is missing. AMI/GMA/ICLR/ICML/USENIX do not optimize attribution, and their hyperparameters are not tuned for this trade-off and AE-APL winning on its own objective is expected. What is missing is an external baseline or simple control that also aims to reduce exposure, for example, regularization toward the benign policy, a sparse attack that intervenes only in critical states, or a matched-exposure baseline that directly weakens the attack. In addition, the capability floor is calibrated from the ICLR baseline (i.e., J_floor), so matched strength depends on the strength of one particular baseline. A hard-threshold gate may also cause 'attack weakens -> gate closes -> attack strengthens' oscillations during training, yet the paper only reports an offline sweep over ρ and does not analyze training stability or sensitivity of the final policy to J_floor. Without these controls, it is difficult to determine whether AE-APL's Pareto frontier is better than simply attacking less or behaving more like the benign policy.
Questions for authors
Q1. Does the paper claim that reducing attribution exposure is equal to reducing true responsibility? If not, can the conclusion be stated explicitly as 'evading a family of post-hoc attribution methods', with a discussion of how an auditor aware of this attack (e.g., multi-step counterfactual, causal responsibility, or an auditor designed to detect blame-spreading) would fare? Q2. The continuous-action counterfactual uses batch-mean interpolation, which is not semantically a feasible benign action. Could this systematically over- or under-estimate c_i? Can you use the benign policy (available only at training time) as an oracle reference to bound the surrogate error? Q3. J_floor depends on the ICLR baseline. If a stronger or weaker reference attack were used, would the conclusions be stable? Does the hard-threshold gate cause training oscillations? Q4. In Table 12, both attacker and benign raw harm decrease while the team still fails. Please provide a mechanistic explanation and validate that 'true responsibility' indeed decreases using a non-Q-based causal/environment-level attribution.