What makes data poisoning challenging to detect compared with traditional data loss?

• A. Data poisoning is harder to detect than traditional data loss because malicious data can be injected at many stages.
• B. Data poisoning is easier to detect than data loss.
• C. Data poisoning only affects training data, making detection straightforward.
• D. Data poisoning is a theoretical risk with minimal real-world impact.

Answer: (A)

Data poisoning is hard to detect because malicious data can be introduced at many points in the data pipeline—from collection and labeling to preprocessing and model updates—and still look like normal, legitimate data. This means standard integrity checks that flag missing or corrupted records aren’t enough, since the data isn’t gone or obviously altered in a noticeable way. Instead, the attacker subtly changes the training signal, sometimes distributing small amounts of poisoned data across many samples or embedding backdoors that only trigger under specific conditions. As a result, the model learns harmful patterns while the dataset’s surface properties still appear intact, making the attack blend in with ordinary data. Detecting this requires deeper scrutiny of data provenance, robust training methods, and targeted testing of model behavior rather than relying on traditional data loss indicators.