Last but not least, came Tekgul & Asokan's "FLARE: Fingerprinting Deep Reinforcement Learning Agents using Universal Adversarial Masks" which is robust to model modification attacks. (https://www.acsac.org/2023/program/final/s264.html) 4/4
#MachineLearningSecurity #DeepReinforcementLearning #SecurityInAI