A model producing contradictory outputs based on highly similar inputs MOST likely indicates the presence...

The AAISM study framework describes evasion attacks as attempts to manipulate or probe a trained model during inference by using crafted inputs that appear normal but cause the system to generate inconsistent or erroneous outputs. Contradictory results from nearly identical queries are a typical symptom of evasion, as the attacker is probing decision boundaries to find weaknesses. Poisoning attacks occur during training, not inference, while membership inference relates to exposing whether data was part of the training set, and model exfiltration involves extracting proprietary parameters or architecture. The clearest indication of contradictory outputs from similar queries therefore aligns directly with the definition of evasion attacks in AAISM materials.