// THREAT DETECTION AND DATA PRIVACY TERM
Data Poisoning
Data poisoning is an attack where a malicious actor intentionally corrupts the data used to train a machine learning model. This causes the model to learn the wrong things, leading to inaccurate or biased results once it's deployed.
TECHNICAL DEFINITION
Data poisoning is a machine learning security attack where an adversary manipulates or injects malicious samples into a model's training dataset to compromise its learning process. This adversarial contamination degrades model performance, reduces accuracy, and can introduce targeted backdoors or biases, affecting classifications and predictions during inference.
BACKGROUND
Adversarial machine learning is the study of the attacks on machine learning algorithms, and of the defenses against such attacks.
READ MORE ON WIKIPEDIASYNONYMS & ALIASES
- adversarial contamination
- dataset poisoning
- training data attack
- model poisoning
- data manipulation attack
- input poisoning
USAGE NOTE
This attack targets the integrity of the model during its training phase, which can be very difficult to detect compared to attacks on a live system.
DEVELOPERS
Organizations developing technology related to Data Poisoning.
A startup focused on AI security that provides a platform to test, validate, and protect machine learning models from vulnerabilities, including data poisoning and adversarial attacks.
An AI security company that develops a Machine Learning Security (MLSec) platform designed to detect and respond to adversarial attacks against machine learning models, including data poisoning techniques.
A not-for-profit organization managing federally funded research and development centers (FFRDCs). They developed the Adversarial Threat Landscape for Artificial-Intelligence Systems (ATLAS) framework, which catalogues and analyzes attacks like data poisoning to build defenses.
The Defense Advanced Research Projects Agency, a research and development agency of the U.S. Department of Defense. DARPA's GARD (Guaranteeing AI Robustness against Deception) program specifically funds the development of defenses against data poisoning and other adversarial ML attacks.
The research and development division for IBM. They actively research and publish on 'Trusted AI,' developing novel algorithms and toolkits to detect and mitigate data poisoning attacks on machine learning models.
A major aerospace and defense technology company that develops AI-enabled systems for military applications. Their work includes research and development into creating trusted and resilient AI that can resist adversarial manipulation, including data poisoning.
The research subsidiary of Microsoft. It has dedicated teams working on 'Responsible AI' and 'Trustworthy Machine Learning,' which includes building frameworks and defenses to secure AI systems from data poisoning and other adversarial threats.
The corporate research lab for Bosch, focusing on foundational AI research. They publish studies and develop methods for creating robust machine learning models that are resilient to data poisoning, particularly for safety-critical applications like autonomous driving.