Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Stealthy Imitation: Reward-guided Environment-free Policy Stealing
Authors: Zhixiong Zhuang, Maria-Irina Nicolae, Mario Fritz
ICML 2024 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5. Experiments. This section presents our empirical results for Stealthy Imitation. We discuss the experimental setup (Section 5.1), followed by a comparison of our proposed method to baselines (Section 5.2) and analyses and ablation studies (Section 5.3). |
| Researcher Affiliation | Collaboration | 1Graduate School of Computer Science, Saarland University, Saarbr ucken, Germany 2Bosch Center for Artificial Intelligence, Robert Bosch Gmb H, Renningen, Germany 3CISPA Helmholtz Center for Information Security, Saarbr ucken, Germany. |
| Pseudocode | Yes | Algorithm 1 Stealthy Imitation |
| Open Source Code | No | 1The project page is at https://zhixiongzh.github.io/stealthyimitation. |
| Open Datasets | Yes | We demonstrate our method on three continuous control tasks from Mujoco (Todorov et al., 2012): Hopper, Walker2D, and Half Cheetah. |
| Dataset Splits | Yes | The transfer dataset Dv described below is split into training and validation for use in the subsequent method steps |
| Hardware Specification | Yes | All experiments were conducted on a single NVIDIA Ge Force RTX 2080 Ti GPU. |
| Software Dependencies | No | The victim policies are trained using the Ding repository (engine Contributors, 2021), a reputable source for Py Torch-based RL implementations (Paszke et al., 2017). |
| Experiment Setup | Yes | We set the reserved training budget Br = 10^6 and the base query budget bv = 10^5. Both πa and πe share the same architecture and are trained for one epoch per iteration. We use the Adam optimizer (Kingma & Ba, 2015) with a learning rate of η = 10^-3 and batch size of 1024. The final training employs early stopping with a patience of 20 epochs for 2000 total epochs. The reward model ˆR is a two-layer fully-connected network (256 hidden neurons, tanh and sigmoid activations). ˆR is trained with a learning rate of 0.001 for 100 steps. |