From the course: CompTIA SecAI+ (CY0-001) Cert Prep

Unlock this course with a free trial

Join today to access over 25,600 courses taught by industry experts.

Model poisoning

Model poisoning

From the course: CompTIA SecAI+ (CY0-001) Cert Prep

Start my 1-month free trial Buy for my team

Model poisoning

“

Model poisoning targets the model itself. Instead of changing the data that teaches the system, the attacker changes the learn parameters. That direct manipulation alters how the model behaves, even when the training data looks clean. Model poisoning takes advantage of a common AI software development pattern. Many teams start with a pre-trained foundation model, adapt it for a new purpose, and then deploy it. These foundation models, such as the GPT models from OpenAI, the Claude models from Anthropic, and the Gemini models from Google, provide an important starting point for AI applications. When a developer adopts one of these models, they adopt the parameters that the model learned during pre-training. Attackers focus on that adoption step. If an attacker persuades a developer to start from a maliciously designed foundation model, the attacker gains influence over every downstream system that relies on it. The tainted model can look capable and helpful. Standard tests can show…

Contents