Dropout is a powerful regularization technique that successfully mitigates overfitting across a wide range of neural network architectures and applications, from computer vision to natural language processing and more. Dropout consists of dropping (i.e., zeroing) a random fraction of the layer’s nodes and then scaling the weights according to the complement probability of the dropout rate. This paper proposes an enhanced weight scaling and uses a simulation study over two benchmark datasets to demonstrate its improved accuracy over the standard approach.

Network anomaly detection is a core research problem. In a previous paper, we proposed the concept of "autoencoder losses transfer learning." This approach normalizes autoencoder losses in different model deployments, and uses a global detection model to detect and classify threats in a generalized way that is agnostic to the specific network deployment, allowing to learn from significantly d ifferent n etworks. W hile t he previous paper provided initial simulations and evaluation results, this paper will present new results of an extensive empirical study of the proposed approach over additional datasets.

A classic long-term challenge is detecting anomalies in computer networks. Despite almost every kind of model architecture, previous works focused on analyzing the complete network traffic, which is becoming less applicable for large networks due to processing overheads. This poster presents a work in progress that transforms the computer network into a graph neural network (GNN) by analyzing only a small fraction of network traffic and detecting anomalies.

In recent years there has been a growing interest in developing "streaming algorithms" for efficient processing and querying of continuous data streams. These algorithms seek to provide accurate results while minimizing the required storage and the processing time, at the price of a small inaccuracy in their output. A fundamental query of interest is the intersection size of two big data streams. This problem arises in many different application areas, such as network monitoring, database systems, data integration and information retrieval. In this paper we develop a new algorithm for this problem, based on the Maximum Likelihood (ML) method. We show that this algorithm outperforms all known schemes in terms of the estimation's quality (lower variance) and that it asymptotically achieves the optimal variance.