Data Poisoning: a danger for Artificial Intelligences

The concept of “data poisoning” poses a threat to AI. This risk can be mitigated, but continues to be a significant topic for cybersecurity and AI-enabled technological progress.

Data Poisoning and its consequences

Data Poisoning is the attack on AI and machine learning by manipulating or injecting altered data into the datasets used to train Machine Learning models. This leads to two catastrophic consequences: the drastic reduction in the reliability of the models and the possibility for attackers to introduce backdoors that allow them to influence the models to their advantage.

Reduce the risk of Data Poisoning

While it is difficult to completely eliminate the risk of Data Poisoning, it is possible to reduce its incidence. However, the data organizations need often comes from sources outside the corporate network, making it nearly impossible to avoid the attack entirely. It is therefore necessary to adopt strategies to limit risk, such as access controls and stricter identification policies.

Defense and awareness techniques

To effectively mitigate the risk of Data Poisoning, it is essential to increase threat awareness and take appropriate defense measures. This includes reducing the attack surface by using firewalls, applying security patches in a timely manner, monitoring network traffic, and preparing an incident response plan. It is also important to also consider physical security to prevent data poisoning within company premises.