On the Steganographic Capacity of Selected Learning Models