Research map / Adversarial robustness and security / Attacks on text and graphs
Budget-Efficient Attacks and Robustness Training for Cooperative MARL
junyong jiang, Xin Yuan, Longhe Lin, Songze Li, Lu Dong
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use BHEA-AT instead of ROMANCE or WALL for adversarial training | QMIX victims · SMAC MMM under BHEA attack | test win rate (%): 80.2 / 63.1 -> 99.2 | Improved |
| adversarial training with BHEA-AT instead of vanilla training | QMIX · SMAC MMM under EGA attack | test win rate (%): 53.1 -> 98.1 | Improved |
| adversarial training with BHEA-AT | QMIX · SMAC natural execution (no attack) | test win rate (%): 99.0 -> 99.4 on MMM | No clear effect |
| BHEA-AT instead of WALL under EGA attack on 8m | QMIX · SMAC 8m under EGA attack | test win rate (%): 77.1 -> 78.3 | No clear effect |
| BHEA-AT instead of ROMANCE under EGA attack on 2s3z | QMIX · SMAC 2s3z under EGA attack | test win rate (%): 94.2 -> 95.3 | No clear effect |
| use BHEA-AT instead of ROMANCE for training cost | adversarial training · SMAC MMM | training time (hours): 30h -> 18h | Improved |
| use BHEA-AT instead of WALL for training cost | adversarial training · SMAC MMM | training time (hours): 47h -> 18h | Improved |
| replace learned victim selector with uniform random selection (-Selector) (ablation) | BHEA-AT · SMAC 1c3s5z and MMM | win rate under attack | Hurt |
| replace learned action attacker with random actions (-Attacker) (ablation) | BHEA-AT · SMAC 1c3s5z and MMM | win rate under attack | Hurt |
| replace learned action attacker with lowest-Q action (-Attacker + WorstQ) (ablation) | BHEA-AT · SMAC 1c3s5z and MMM | win rate under attack | Mixed |
| use full BHEA instead of ablated selector or attacker variants (ablation) | BHEA attack · SMAC 1c3s5z and MMM | attack strength (victim win rate) | Improved |
| increase per-timestep attack capacity K (ablation) | BHEA · SMAC 1c3s5z (B=4), MMM (B=3) | victim win rate | Improved |
Closest papers by meaning
- Rethinking Adversarial Policies: A Generalized Attack Formulation and Provable Defense in RL (ICLR 2024 · similarity 0.88)
- Wolfpack Adversarial Attack for Robust Multi-Agent Reinforcement Learning (ICML 2025 · similarity 0.88)
- Byzantine Robust Cooperative Multi-Agent Reinforcement Learning as a Bayesian Game (ICLR 2024 · similarity 0.87)
- Robust and Diverse Multi-Agent Learning via Rational Policy Gradient (NeurIPS 2025 · similarity 0.87)
- AdvEvo-MARL: Shaping Internalized Safety through Adversarial Co-Evolution in Multi-Agent Reinforcement Learning (ICML 2026 · similarity 0.86)
- Beyond Worst-case Attacks: Robust RL with Adaptive Defense via Non-dominated Policies (ICLR 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
