System and method for protection and detection of adversarial attacks against a classifier

A system for detection of adversarial attacks against a classifier, the system comprising:

• a non-transitory memory device, wherein modules of instruction code are stored, and at least one processor associated with the memory device, and configured to execute the modules of instruction code, whereupon execution of said modules of instruction code, the at least one processor is further configured to:
• receive a first data element comprising at least one adversarial data perturbation;
• apply the classifier on the first data element, to produce a first classification of the data element;
• generate, by an autoencoder, a reconstructed, second data element based on the first data element, the second data element is devoid of the at least one adversarial data perturbation;
• apply the classifier on the second data element, to produce a second classification of the data element;
• compare, by a comparator, the first and second classifications;
• and
• determine, based on the comparison, that an adversarial attack has occurred when the first classification and the second classification are different, and that an adversarial attack has not occurred when the first classification and the second classification are equivalent.