Supervised contrastive learning over prototype-label embeddings for network intrusion detection

Contrastive learning makes it possible to establish similarities between samples by comparing their distances in an intermediate representation space (embedding space) and using loss functions designed attract/repel similar/dissimilar samples. The distance comparison is based exclusively on the sample features. We propose a novel contrastive scheme including labels same embedding as features performing this shared space. Following idea, should be close its ground-truth (positive) label away from other (negative labels). This allows implement supervised classification learning. Each embedded will assume role of class prototype space, with that share gathering around it. aim separate prototypes while minimizing each same-class A set proposed objective. Loss minimization drive allocation associated training prediction architectures are analyzed detail, along different strategies for separation. drastically reduces number pair-wise comparisons, thus improving model performance. In order further reduce initial extended replacing negative best single representative: either nearest or centroid cluster labels. idea creates new subset models which detail. outputs (in prototypes. These can used perform (minimum label), dimensionality reduction (using embeddings instead original features) data visualization (with 2 3D embeddings). Although generic, application performance evaluation done here network intrusion detection, characterized noisy unbalanced challenging various types attacks. Empirical results applied detection presented detail two well-known datasets, thorough clustering metrics included.