Information Bottleneck under Perfect Privacy
The paper adds exact statistical privacy to the information bottleneck objective and gives an ADMM solver for that constrained problem.
The authors focus on the active-rate regime, where the representation-rate limit is binding and directly caps utility. Their setup requires the learned representation to preserve utility-relevant information while remaining independent of a sensitive variable. They treat that independence as an additional optimization constraint, not as a side condition. They prove global convergence under regularity assumptions, describe rates via the Kurdyka-Lojasiewicz exponent, and extend the result to inexact block updates. ArXiv · AI/CL/LG's note
The authors focus on the active-rate regime, where the representation-rate limit is binding and directly caps utility. Their setup requires the learned representation to preserve utility-relevant information while remaining independent of a sensitive variable. They treat that independence as an additional optimization constraint, not as a side condition. They prove global convergence under regularity assumptions, describe rates via the Kurdyka-Lojasiewicz exponent, and extend the result to inexact block updates. ArXiv · AI/CL/LG's note
score 4