Differentially Private Label Protection in Split Learning

This is an academic paper (arXiv preprint) with a reference implementation addressing a specific vulnerability in split learning architectures. The defensibility score reflects: (1) Zero stars and forks suggest minimal adoption or visibility in the open-source ecosystem; (2) Academic provenance (paper published ~5 years ago with no subsequent versioning signals) indicates this is research-grade work, not a production system; (3) The novelty is a sound one—combining differential privacy mechanisms with split learning's label-sharing phase—but the contribution is incremental within the federated/privacy-preserving ML landscape, which already has substantial prior work (FedProx, secure aggregation, etc.). Platform domination risk is medium because large cloud providers (AWS SageMaker, Google Vertex AI, Azure ML) are actively investing in federated learning and privacy-preserving training frameworks; they could absorb this technique as a built-in defense. Market consolidation risk is low because the split learning market itself is nascent and fragmented—no dominant incumbent has formed yet, so acquisition pressure is minimal. Displacement horizon is 3+ years because split learning adoption remains research-focused; real production deployments are limited, and privacy-preserving distributed training is still maturing. The integration surface is limited to reference code and the algorithm itself, making it primarily useful to researchers and practitioners implementing split learning from scratch. This is not a library, service, or framework competitors would fork or extend; it's a defense mechanism that would be integrated into larger federated learning platforms. No evidence of real-world deployment, community adoption, or ecosystem lock-in. The 5 forks likely represent academic citations or students reproducing experiments, not production usage.