Secure Learning and Mining in Adversarial Environments [Extended Abstract]

Machine learning and data mining have become ubiquitous tools in modern computing applications and large enterprise systems benefit from its adaptability and intelligent ability to infer patterns that can be used for prediction or decision-making. Great success has been achieved by applying machine learning and data mining to the security settings for large dataset, such as in intrusion detection, virus detection, biometric identity recognition, and spam filtering. However, the strengths of the learning systems, such as the adaptability and ability to infer patterns, can also become their vulnerabilities when there are adversarial manipulations during the learning and predicting process. Considering the fact that the traditional learning strategies could potentially introduce security faults into the learning systems, robust machine learning techniques against the sophisticated adversaries need to be studied, which is referred to as secure learning and mining through this abstract. Based on the goal of secure learning and mining, I aim to analyze the behavior of learning systems in adversarial environments by studying different kinds of attacks against the learning systems. Then design robust learning algorithms to counter the corresponding malicious behaviors based on the evaluation and prediction of the adversaries' goal and capabilities. The interactions between the defender and attackers are modeled as different forms of games, therefore game theoretic analysis are applied to evaluate and predict the constraints for both participants to deal with the real world large dataset.