Transferable Environment Poisoning: Training-time Attack on Reinforcement Learning
Bibliographic record
Abstract
Studying adversarial attacks on Reinforcement Learning (RL) agents has become a key aspect of developing robust, RL-based solutions. Test-time attacks, which target the post-learning performance of an RL agent's policy, have been well studied in both white- and black-box settings. More recently, however, state-of-the-art works have shifted to investigate training-time attacks on RL agents, i.e., forcing the learning process towards a target policy designed by the attacker. Alas, these SOTA works continue to rely on white-box settings and/or use a reward-poisoning approach. In contrast, this paper studies environment-dynamics poisoning attacks at training time. Furthermore, while environment-dynamics poisoning presumes a transfer-learning capable agent, it also allows us to expand our approach to black-box attacks. Our overall framework, inspired by hierarchical RL, seeks the minimal environment-dynamics manipulation that will prompt the momentary policy of the agent to change in a desired manner. We show the attack efficiency by comparing it with the reward-poisoning approach, and empirically demonstrate the transferability of the environment-poisoning attack strategy. Finally, we seek to exploit the transferability of the attack strategy to handle black-box settings.
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 | 0.002 | 0.008 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.002 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.001 | 0.002 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.002 | 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".