Generating Malicious Demonstration Policies to Exploit Vulnerabilities in Inverse Reinforcement Learning
Bibliographic record
Abstract
Reinforcement Learning (RL) algorithms depend on well-defined reward functions for policy optimization. Designing such functions is a complex task, even for domain experts. However, valid task demonstrations can still be collected from moderately skilled individuals. This motivates the use of methods such as Imitation Learning (IL) and Inverse Reinforcement Learning (IRL), where an expert provides demonstrations, allowing the algorithm to infer a policy or reward function that aligns with observed behavior. A common assumption in IRL is that demonstrations come from highly skilled experts. While some studies have explored the impact of suboptimal demonstrators, the influence of intentionally malicious demonstrations remains underexplored. This study introduces an adversarial demonstrator framework that systematically perturbs a subset of demonstrations to manipulate the reward function inferred by an IRL algorithm. Additionally, it quantifies the impact of such adversarial manipulations on the learned policy. Our results show that simply altering 10% of the demonstrations can lead the IRL algorithm to learn a faulty reward function, ultimately degrading the performance of the trained policy by up to three times the effect of adding 10% random trajectories to the result.
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 | 0.018 |
| 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 | 0.002 |
| 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".