Stealthy Attacks on Multi-Agent Reinforcement Learning in Mobile Cyber-Physical Systems
Bibliographic record
Abstract
Due to their mobility, real-time requirements, energy limitations, and safety considerations, the complexities involved in Mobile Cyber-Physical Systems (MCPSs) surpass those of traditional computing systems. To address these challenges, the use of multi-agent reinforcement learning (MARL) algorithms is gaining significance in the field of MCPS. MARL enables precise, instantaneous, and coordinated decision-making to maximize cumulative rewards through systematic trial and error, even in unfamiliar environments. While MARL algorithms can effectively learn scalable and efficient control policies for MCPSs, their resilience against security and safety attacks has not been thoroughly explored, severely limiting their real-world applications. This paper investigates the robustness of MARL-based MCPS against stealthy adversarial attacks which involve targeting and manipulating a specific mobile node in order to generate deceptive observations that adversely affect the behavior of other MCPS nodes. We adopt the FGSM (Fast Gradient Sign Method) adversarial example technique from deep learning to incorporate a detection evasion mechanism as a new stealth feature. The objective is to entice the compromised node to adopt an adversarial policy that deviates the activations of policy networks in its cooperative nodes from the expected distribution, while evading detection. We empirically demonstrate the susceptibility of MARL algorithms commonly employed in MCPSs, namely MADDPG, to our proposed attack strategies. The evaluation is conducted in three MCPSs, considering both white and black-box settings. By targeting a single node, our attacks have a significantly detrimental impact on the overall performance of the MCPS, resulting in a minimum reduction of 33% and a maximum reduction of 89.6% in the system's overall reward, with an evasion rate ranging from 16% to 36%.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.005 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.001 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.001 | 0.001 |
| Insufficient payload (model declined to judge) | 0.001 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".