2017 Day 1 Nick Savvides: Combating Advanced Cybersecurity Threats with AI

Nick Savvides distinguishes between true artificial intelligence and machine learning, explaining that AI requires the ability to perceive, perform cost-benefit analysis, and make autonomous decisions (exemplified by KITT from Knight Rider), whereas machine learning focuses on pattern recognition and self-development. He traces ML history from the 1950s to 2006, marking 2006 as the "democratization of machine learning" when computational power became affordable enough to run deep learning algorithms. The Netflix Prize exemplified this shift and drove significant ML development.

Savvides contrasts traditional programming (source code → compiled program + data → output) with machine learning (algorithm + training data → model → output). He explains supervised learning (with labeled "good" and "bad" examples) versus unsupervised learning (unlabeled data), noting that cybersecurity predominantly uses unsupervised approaches.

He identifies three primary ML applications in cybersecurity: threat detection, user behavior analysis (detecting anomalies through behavioral clustering), and anomaly detection (identifying unexpected patterns). ML is deployed across the security stack—from endpoints to networks to security operations centers—though different models must be tailored for resource-constrained endpoints versus well-resourced network infrastructure.

At the endpoint level, ML can protect disconnected environments (models don't require constant signature updates), detect unknown threats including non-portable executables, provide feedback to other detection systems, and extract important features for discovering new indicators of compromise. At the network level, ML enables more complex models, reinvents existing data sources (such as repurposing IPS signatures), auto-prioritizes human investigations, discovers unknown breaches, and extracts secondary indicators of compromise.

However, ML systems face adversarial attacks: model extraction (reverse-engineering models through API queries), poisoning (corrupting training data to degrade model performance), and evasion (crafting inputs to evade detection). Attackers can weaponize ML themselves—for example, using ML-powered chatbots to impersonate executives in business email compromise attacks, a capability Microsoft demonstrated six years prior.

Savvides predicts ML will evolve from detecting threats to predicting attacks, but expresses critical concern: if organizations cannot understand why a model makes its recommendations, they risk losing control of their security systems. He concludes that every organization must consider applying machine learning across operations, emphasizing it is a current, accessible technology rather than a future prospect.