Machine Learning Framework for Security Applications

Speaker: Paul Irofti (University of Bucharest).

Machine learning helps us tackle large and apparently intractable optimization problems. Even though neural networks are by far the most popular choice in the field, we focus on dictionary learning (DL) for sparse representations (SR) instead. Our preference is motivated by the much simpler model that provides faster methods with a solid theoretical background, understanding and interpretability.

In fact it has been recently shown that the forward pass inside neural networks is equivalent to performing sparse representation. Thus performing dictionary learning can be interpreted as a backward pass on a much simpler and smaller model. This relaxation comes with a small performance hit in exchange for the large reduction in algorithm complexity.

Our talk will focus on adapting DL to Big Data conditions, DL classification and the problem of malware identification,  nomaly detection, online DL and Internet of Things applications.