Privacy Against Agnostic Inference Attacks in Vertical Federated Learning

The project is a code accompaniment to an academic paper (arXiv:2302.05545). From a competitive standpoint, it scores low on defensibility (2) because it is a reference implementation with no community traction (0 stars), focusing on a limited model type (Logistic Regression). While the research addresses a valid problem in Vertical Federated Learning (VFL)—where the party with the labels can exploit the gradient updates to guess the other party's data—the code itself lacks the robust packaging or framework-level integration (like FATE or PySyft) required for production use. Frontier labs like OpenAI or Google are unlikely to target this specific niche of VFL directly, as their focus remains on Horizontal FL and LLM alignment; however, the algorithmic contribution could be easily absorbed into larger privacy-preserving ML frameworks. The displacement horizon is relatively short (1-2 years) because academic defenses in FL are rapidly evolved or replaced by more generalized differential privacy or homomorphic encryption techniques.