Machine learning is rapidly gaining popularity in the security space. Many vendors and security professionals are touting this new technology as the ultimate malware defense. While evidence from both research and practice validates the improved efficacy of machine learning-based approaches, their drawbacks are rarely discussed.
In this talk, we will demonstrate, from an attacker's perspective, how commonly deployed machine learning defenses can be defeated. We then step back and examine how existing systemic issues in the network security industry allow this to occur, and begin the discussion with the community about these issues. Finally, we propose a solution that uses novel data sourcing techniques to address these problems.