Learning Constant-Depth Circuits in Malicious Noise Models