Data Free Black Box Attack

Data-free black-box attacks aim to compromise machine learning models without access to their training data, focusing instead on manipulating model outputs or internal parameters. Current research explores techniques like optimal transport for model fusion to mitigate these attacks, dynamically adapting substitute model architectures for improved attack effectiveness, and generating synthetic data to poison federated learning systems. These attacks pose a significant threat to the security and reliability of deployed machine learning models, driving research into robust defenses and highlighting the need for more secure model training and deployment strategies.