The Unspoken Problems With Machine Learning in Security

Machine Learning is the hottest buzzword. Everybody loves it, everybody sells it. But why is it that while fields such as Computer Vision or Natural Language Processing have stellar achievements, with new record-breaking models published every other week, the Cybersecurity industry staggers behind?

Are Anomaly Detection algorithms – so well beloved for the prevention of attacks and of fraud - really suitable for those intended purposes? What price do we pay for keeping things hushed? Where do our huge datasets fail us? And how, once we spot these issues, might we try and solve them?

In this talk I will go over several points that hold us back, among them: our rapidly changing input data, and who’s to blame for it; the known issues of imbalanced & untagged datasets, and why our solutions for them are insufficient; and, finally, the biggest culprit: the confidential nature of our field, and how it keeps us from being great.

The unspoken problems with Machine Learning in security: let’s talk about them.

The Quick & Dirty AI Startup

You just founded your AI startup - and now it’s time to build your algorithm. But, beware: when building an AI-based product from scratch, it’s easy to get lost in the details, investing effort where it’s not needed.

Join me to hear of simple ways to save R&D time without compromising on product quality, gathered from my experience consulting for multiple early-stage startups. We’ll discuss what’s important to focus on when building your algorithms, where you want to invest your time, and what corners you should cut. We will also talk about the right time to hire an in-house data scientist, and how to make do before that.

If you’re thinking of starting your own AI startup, already have one, or are a sole data scientist in a brand-new venture - this talk is for you.

The Unspoken Problems With Machine Learning in Security

Machine Learning is the hottest buzzword. Everybody loves it, everybody sells it. But why is it that while fields such as Computer Vision or Natural Language Processing have stellar achievements, with new record-breaking models published every other week, the Cybersecurity industry staggers behind?

Are Anomaly Detection algorithms – so well beloved for the prevention of attacks and of fraud - really suitable for those intended purposes? What price do we pay for keeping things hushed? Where do our huge datasets fail us? And how, once we spot these issues, might we try and solve them?

In this talk I will go over several points that hold us back, among them: our rapidly changing input data, and who’s to blame for it; the known issues of imbalanced & untagged datasets, and why our solutions for them are insufficient; and, finally, the biggest culprit: the confidential nature of our field, and how it keeps us from being great.

The unspoken problems with Machine Learning in security: let’s talk about them.

AI/ML Street Smarts: Navigating the Sea of Information

The field of AI is a fast-changing one: not a week goes by without some new and exciting development. As professionals in the field, we strive to stay up to date with the cool kids - while not losing focus on our own work.

In this talk I’ll discuss that challenge, the different ways to approach it, and the varied resources one might use to that aim. I will also present common struggles and their not-so-common solutions, and talk about how we might achieve that holy grail: the perfect balance between exploring new topics and our day-to-day work.

10 Must-Know Concepts in AI & Machine Learning

Do you often find yourself in meetings flooded with obscure AI terminology, wondering what it all means? There is hope yet. Designed for tech professionals, this talk introduces 10 must-know concepts in AI/ML, giving you the basic tools to better understand the work around you. Rock the next water cooler chat with your newfound knowledge of the fundamentals of AI and machine learning.

Target audience: tech professionals who are not AI/ML engineers.

Preferred duration: 30-50 minutes.

Choosing the Right Machine Learning Abstraction for your Business Needs

Choosing the right abstraction for your problem - it is a step crucial to the success of every ML project, yet one that is often overlooked. We could spend hours debating whether to use XGBoost or CATboost, yet neglect giving our conscious attention to a more elementary decision: how to model our problem in the first place.
The exact same business case could be modeled, for example, as a classification problem, a clustering one, or even a graph link-predictions task. And as the options vary, so do the considerations for choosing among them, including not only machine learning theory, but also the needs of your business, organizational constraints, and many more.
Join me to discuss the principles of choosing the best ML abstraction for your needs, things to consider and pitfalls to avoid, all through the lens of a real-life case study from my consultancy work.

Outline:
• Self intro (3 minutes)
• What do I talk about when I talk about “choosing the right ML abstraction” (5 minutes)
◦ (Explain concept and how it relates to business strategy)
• Go through a use case (15 minutes total)
• Overview of the business case (5 minutes)
• Modeling alternatives: (7 minutes)
◦ First iteration: go through just ML pros and cons
◦ Then, another iteration – adding business needs pros and cons
• Present EL’s final decision & explain reasoning - 3 min
• Summary & take-home message (2 minutes)
• Q&A (5 minutes)