AI Data Supply Chain Data poisoning: the quietest way to undermine an AI model.

AI Data Supply Chain Data poisoning: the quietest way to undermine an AI model.

When it comes to impacting AI models, forget prompt injection; the real battleground lies upstream. The key is to manipulate the dataset, giving you control over the model.

Here’s a breakdown of how it unfolds:

🔹 Attack surfaces:
• Publicly scraped web data: Plant deceptive information and wait for it to be scraped.
• Crowdsourced/contributed datasets: Conceal mislabels in plain sight.
• Open repositories: Upload a seemingly helpful dataset with a hidden agenda.
• Fine-tuning/RLHF feedback: Act as the "supportive" annotator subtly sabotaging the model.

🔹 Techniques:
• Label flipping: Disguise spam as "ham" to deceive the model.
• Trigger backdoors: Embed a secret word or pixel pattern to manipulate the output.
• Semantic corruption: Introduce falsehoods like "5G causes cancer" until the model accepts them as truth.
• Gradient steering: Shape data to influence the model towards the attacker's objectives.

The risk?
It's not just about having an inaccurate model; it’s about possessing a compromised one. This compromised model willingly carries the attacker's payload into every subsequent application and API.

Data poisoning is akin to cancer within the bloodstream. Once it infiltrates the training set, its impact spreads extensively.

👉 The future of AI security extends beyond model surveillance; it's about safeguarding the data supply chain.