Audit Algorithms

Finds groups of features that can be permuted without changing the output label of predicted samples

Aims at extracting machine learning models in use by remote services.

Explains a blackbox classifier model by sampling around data instances.

Infer inner hyperparameters (eg number of layers, non-linear activation type) of a remote neural network model by analysing its response patterns to certain inputs.

Stealing black-box models (CNNs) knowledge by querying them with random natural images (ImageNet and Microsoft-COCO).

Check if a remote machine learning model is a "leaked" one: through standard API requests to a remote model, extract (or not) a zero-bit watermark, that was inserted to watermark valuable models (eg, large deep neural networks).

Shows the impossibility (with one request) or the difficulty to spot lies on the explanations of a remote AI decision.

Crafts adversarial examples to fool models, in a pure blackbox setup (no gradients, inferred class only).

Parametrize a local recommendation algorithm by imitating the decision of a remote and better trained one.

(arxiv) Infers a link between the Amazon Echo system and the ad targeting algorithm.

(Journal of Information Science) Audits multiple search engines using simulated browsing behavior with virtual agents.

(NeurIPS - best paper) A scheme for auditing differentially private machine learning systems with a single training run.

Relates the difficulty of black-box audits to the capacity of the targeted models, using the Rademacher complexity.

Sequential methods that allows for the continuous monitoring of incoming data from a black-box classifier or regressor.