What is a data poisoning attack and how can you detect it?

• A. Training With Incorrect Labels Without Detection
• B. Data Leakage Through Encryption
• C. Poisoning Inference Phase With Noise
• D. Inserting Malicious Data Into Training; Detection Via Data Provenance, Anomaly Detection In Data, Validation On Hold-Out Clean Data, Robust Training

Answer: (C)

Data poisoning involves injecting malicious or mislabeled examples into the training data to steer the model toward specific, harmful outcomes. Detecting it relies on tracing data provenance to verify sources, spotting anomalies or unusual patterns in the data, and validating the model on a hold-out clean dataset that wasn’t used during training. Robust training practices—such as data sanitization, anomaly-resistant objectives, and reweighting or filtering suspicious samples—help reduce the impact of poisoned data.

The description that captures data poisoning most accurately is the act of inserting malicious data into the training process, with detection anchored in provenance checks, data anomaly detection, validation on clean hold-out data, and robust training methods. The idea of poisoning during the inference phase with noise describes adversarial examples at inference time, which is a different kind of attack.