E178 ‘Interactive and Incremental AI in Security’ with IBM Research’s Seshu Tirupathi and Elizabeth Daly

Today’s guests are Seshu Tirupathi, Research Scientist and Elizabeth Daly, STSM and Research Manager of the Interactive AI Group at IBM Research Ireland. In today’s episode, we are delving into fascinating world of AI and its application in cybersecurity, spotlighting IBM Research’s work as a finalist for Best Application of AI in a Large Enterprise at last month’s AI Awards. IBM’s innovative approach to intertwining interactive AI with incremental machine learning is not just changing the way we think about digital security – it’s redefining it.

IBM’s project stands out due to its intricate blend of interactive incremental machine learning and risk classification. Elizabeth shares how this project emerged from a seamless synergy between different disciplines. The team tackled the challenge by fusing their experiences in interactive AI and incremental learning to create a solution for IBM’s Security Guardium product. This novel approach not only simplified but also made feasible the integration of complex AI systems into everyday cybersecurity tasks.

The technical challenges were immense. Seshu explains that moving from traditional heuristic models to incremental learning models was a pivotal step. Continuous learning, dynamic changes in features and ensuring explainability were core challenges they overcame. The project’s aim of allowing a seamless update to models without starting from scratch was ambitious but necessary to maintain robust security.

Elizabeth and Seshu also delve into the advantages of a dual approach. This enables organisations to better safeguard their sensitive information compared to static security measures. Elizabeth highlights how the project allowed integration of expert feedback directly into their machine learning models, making the approach both dynamic and reliable.

One of the significant breakthroughs of IBM’s system is the real-time tracking capabilities, overcoming limitations faced by traditional security methods. Elizabeth highlights the system’s ability to generate synthetic data inputs into machine learning models, marrying expert knowledge with data-driven AI insights. This not only enhances the model’s performance but ensures it remains relevant in fast-evolving environments.

A cornerstone of IBM’s solution is its focus on explainability. As Seshu explains, understanding decision outputs is crucial for users, particularly in a field as sensitive as security. The integration of interpretable models ensures that users know why certain decisions were made, fostering trust and enabling effective user feedback.

The project leverages IBM’s open-source BRCG algorithm, which propels explainability to new heights by creating logical rule-based surrogates. By distilling machine learning models into easily comprehensible rules, IBM has bridged the gap between complex algorithms and user comprehension.

For IBM, trust and human-centric design remain at the heart of AI development. Elizabeth reflects on the project’s inception, driven by the need for stakeholders to intuitively grasp AI-driven decisions. This transparency not only aids compliance with external policies but enhances user confidence in digital tools.

Seshu wraps up the discussion by touching on future trends. The increasing complexity of IT systems, coupled with large volumes of data, demands scalable and adaptable models. As the digital landscape continues to evolve, IBM’s pioneering work in adaptive and interactive machine learning will undoubtedly play a pivotal role.

In our rapidly digitising world, the importance of robust and adaptable security measures cannot be overstated. IBM’s innovative use of interactive AI and incremental learning is a testament to their commitment to addressing present-day challenges with forward-thinking solutions. As we look ahead, this revolutionary approach to cybersecurity not only sets a benchmark for innovation but offers a glimpse into the future of AI-driven solutions.