HARPO trains an RL-based obfuscation agent by analyzing a user’s browsing profileusing an embedding and then optimizing the reward by inter-acting with a black-box user profiling or ad targeting model