Adversarial Representation Learning with Closed-Form Solvers

Adversarial representation learning aims to learn data representations for a target task while removing unwanted sensitive information at the same time. Existing methods model parameters iteratively through stochastic gradient descent-ascent, which is often unstable and unreliable in practice. To overcome this challenge, we adopt closed-form solvers adversary task. We them as kernel ridge regressors analytically determine an upper-bound on optimal dimensionality of representation. Our solution, dubbed OptNet-ARL, reduces stable one one-shot optimization problem that can be solved reliably efficiently. OptNet-ARL easily generalized case multiple tasks attributes. Numerical experiments, both small large scale datasets, show that, from perspective, exhibits three five times faster convergence. Performance wise, when attributes are dependent, learns offer better trade-off front between (a) utility bias fair classification (b) privacy by mitigating leakage private than existing solutions.