Trade-offs and Guarantees of Adversarial Representation Learning for Information Obfuscation