Algorithmic HUMINT: Deep-Learning AI for Counter-Cartel HUMINT Operations

Artificial intelligence (AI) offers law-enforcement and counterintelligence services a credible path to penetrating and disrupting transnational drug cartels. This article advances a unified framework, Algorithmic HUMINT, that integrates three deep-learning components—convolutional neural networks (CNNs), the You Only Look Once (YOLO) family of object detectors, and the ArcFace additive angular-margin loss—into the classical human-intelligence (HUMINT) tradecraft cycle. The framework treats AI not as a substitute for the case officer, but as a force multiplier that absorbs the data-intensive tasks of identification, surveillance, and target profiling, leaving the inherently human work of recruitment and source handling to trained professionals. We share findings from Project JANUS, a comprehensive AI research effort at Mercyhurst University. The project team built an open-source intelligence (OSINT) gallery of 189 images of 13 senior Sinaloa Cartel figures. The three AI computer vision algorithms were built and tested in Python using TensorFlow/Keras, the Ultralytics YOLO library, and the InsightFace ArcFace implementation. A baseline CNN, built from scratch, hit 96.1% training accuracy after 25 epochs. The YOLOv8 classifier used pre-trained ImageNet weights and reached much higher validation accuracy. Pre-trained ArcFace embeddings produced within-identity cosine similarities of 0.55 to 0.85, while impostor pairs scored only 0.05 to 0.35. That gap matches the model’s reported 99.83% accuracy on Labeled Faces. We conclude that pre-trained angular-margin embeddings are the strongest part of the pipeline for biometric surveillance of cartel high-value targets. We also argue for a strict human-in-the-loop setup to manage bias, civil-liberties risks, and adversarial threats.