HIPAA Tip: Artificial Intelligence Poisoning Attacks
Artificial Intelligence or AI Poisoning Attacks take two significant forms: Data Poisoning and Model Poisoning.
Data Poisoning is classified into two categories: targeted data poisoning and non-targeted data poisoning. Targeted attacks occur when an adversary attempts to manipulate the model’s behavior with respect to a specific situation. An example would be a cybercriminal introducing poisoned data to a model designed to detect malware, causing it to miss certain threats. Non-targeted attacks would weaken the model’s ability to process data correctly; think of small, imperceptible flaws in the model’s decision-making process.
AI Model poisoning is a cyberattack that manipulates an AI’s training data to create vulnerabilities, induce bias, or trigger backdoors. Attackers insert tainted samples into training sets, alter labels, or hijack data sources, allowing the model to learn faulty associations. Models act normally until a specific “trigger” word or signal is provided, prompting malicious behavior.
How to look for Poisoning:
- Audit Data Provenance and Integrity regularly, using tools to track changes and verify data origins.
- Use Anomaly Detection Algorithms to identify outliers in the training data; i.e., mislabeled instances.
- Perform Model Behavior Monitoring for deviations from established baselines.
Run Adversarial Red Teaming attacks by creating poisoned inputs to test how the models behave and check for hidden back doors.
Compliance Is Ongoing
HIPAA compliance isn’t a one-time checklist. It’s ongoing, programmatic in nature, and requires demonstrated reasonable diligence to stay in good standing with the U.S. Department of Health and Human Services (HHS) Office for Civil Rights (OCR). Anatomy IT can provide you peace of mind with our expert HIPAA compliance services. To learn more, contact us here.
