privora-graph

Security & Privacy Policy

Architectural Controls in privora-graph

  1. Mathematically Certified Per-Example Sensitivity (DP-SGD): Every individual sample gradient is bounded strictly to $L_2 \le C$ prior to aggregation, and cumulative $(\varepsilon, \delta)$ privacy loss is accounted via exact log-space Rényi divergence without asymptotic approximations.
  2. Immutable Bi-Temporal Audit Trail: Historical graph assertions are never mutated or deleted in place; superseding facts close tx_superseded and append a new tx_recorded edge, guaranteeing point-in-time forensic reproducibility.
  3. Byzantine Gradient Filtering: Multi-Krum and Smoothed Weiszfeld RFA bound adversarial silo influence up to breakdown point $f < K/2$.

Reporting a Vulnerability

Please report security or privacy issues responsibly to sendemailtonagarjuna@gmail.com.