Reputational Damage

Reputational damage happens due to bad publicity. For example, when an autonomous vehicle (AV) is fooled into not recognizing a stop sign or changing lanes to go into opposite traffic, the organization that developed the AV receives significant bad publicity. Such attacks can also significantly decrease the reliability or availability of the AV. When the attacks compromise the confidentiality or privacy of sensitive data managed by ML system, the event can register as a data breach and the organization may have to notify individuals whose data were compromised. If ML system is embedded in a physical artifact such as a robot or an AV, attackers can also have the ML system take actions that result in physical harm to people or property. The organization can also incur financial damages due to these harms. We define the magnitude of harms caused by an adversarial attack on ML system to the victimized organization as the total number of unique harm types caused to the victimized organization.