Detection of privacy attacks on machine learning models
A plurality of queries are input into an artificial intelligence (AI) model. The AI model is made up of a plurality of layers including an input layer, an output layer, and at least one intermediate layer between the input layer and the output layer. Each intermediate layer, during inference, can output a plurality of activations. Thereafter, for each query, activations are intercepted from at least one of the intermediate layers. It is then determined whether a distribution of the intercepted activations across the queries indicates that the queries seek to cause the AI model to behave in an undesired manner by conducting a distance-based similarity analysis between the intercepted activations and reference activations. Data characterizing such determination is then provided to a consuming application or process. Related apparatus, systems, techniques and articles are also described.
1. A computer-implemented method comprising:
inputting each of a plurality of queries into an artificial intelligence (AI) model comprising a plurality of layers including an input layer, an output layer, and at least one intermediate layer between the input layer and the output layer, each intermediate layer, during inference, outputting a plurality of activations;
intercepting, for each query, activations from at least one of the intermediate layers;
determining whether a distribution of the intercepted activations across the queries indicates that the queries seek to cause the AI model to behave in an undesired manner by conducting a distance-based similarity analysis between the intercepted activations and reference activations, the distance-based similarity analysis comprising:
generating a first metric which characterizes overall divergence between the intercepted activations and the reference activations; and
generating a second metric which groups queries by class predictions and calculates a distance for each group relative to reference activations of a same class; and
providing data characterizing the determining to a consuming application or process.
2. The method of claim 1 , wherein the AI model comprises one or more neural networks.
3. The method of claim 1 , wherein the behaving in an undesired manner comprises leaking privacy information originating from a training data set of the AI model.
4. The method of claim 1 , wherein the reference activations comprise distributions of activations empirically derived using the AI model.
5. The method of claim 4 , wherein the reference activations are generated from queries known to be benign.
6. The method of claim 4 , wherein the reference activations are generated from queries known to be malicious.
7. The method of claim 6 , wherein the reference activations are further generated from queries known to be benign such that the similarity analysis is conducted relative to both reference activations known to be benign and reference activations known to be malicious.
8. The method of claim 1 , wherein the reference activations comprise approximated distributions of activations.
9. The method of claim 8 , wherein the approximated distribution of layers are generated using a batch normalization layer initialization strategy.
10. The method of claim 8 , wherein the approximated distribution of layers are generated using public data having a distribution similar to a training data set used to train the AI model.
11. The method of claim 8 , wherein the approximated distribution of layers are generated by initializing activations with random noise.
12. The method of claim 8 further comprising:
initializing activations approximating a training data set used by the AI model; and
optimizing the initialized activations to generate outputs having a likelihood above a pre-defined level.
13. The method of claim 12 , wherein the optimizing is performed using one or more adversarial data generation techniques.
14. The method of claim 1 , wherein the query comprises a privacy attack and the method further comprises:
identifying a training point in a training data set used to train the AI model as being targeted by the privacy attack.
15. The method of claim 14 further comprising:
identifying, for each training point in the training data set, k-nearest neighboring activations associated with an activation for the training point; and
computing, for each training point in the training data set, an average distance from the activation for the training point to the identified k-nearest neighboring activations;
determining that at least one of the queries includes an activation within the average distance for a particular training point; and
flagging the particular training point as being targeted by the privacy attack.
16. The method of claim 1 , wherein the first metric is based on a Fréchet distance and the second metric is based on a class-conditioned Fréchet distance.
17. The method of claim 16 , wherein the AI model comprises a large language model, each of the queries comprises a prompt and the method further comprises:
identifying, for each prompt, a function call name associated with the prompt; and
wherein classes used by the class-conditioned Fréchet distance are based on the identified function call names.
18. The method of claim 15 , wherein at least one of the first threshold and the second threshold is calculated by constructing benign and adversarial queries and optimizing for a cost function.
19. The method of claim 15 , wherein at least one of the first threshold and the second threshold is calculated by constructing benign queries and setting the corresponding threshold to be greater than a distance observed with the benign queries.
20. The method of claim 1 , wherein the reference activations comprise surrogate activations having a dimensionality lower than activations associated with the AI model.
21. The method of claim 1 , wherein the consuming application or process initiates at least one remediation action based on the provided data.
22. The method of claim 21 , wherein the at least one remediation action prevents a subsequent query from being input into the AI model.
23. The method of claim 21 , wherein the at least one remediation action flags the query as being malicious.
24. The method of claim 21 , wherein the at least one remediation action modifies a subsequent query received from a requester of the queries to be benign.
25. The method of claim 21 , wherein the at least one remediation action blocks an internet protocol (IP) address of a requester of the queries.
26. A computer-implemented method comprising:
inputting each of a plurality of queries into an artificial intelligence (AI) model comprising a plurality of layers including an input layer, an output layer, and at least one intermediate layer between the input layer and the output layer, each intermediate layer, during inference, outputting a plurality of activations;
intercepting, for each query, activations from at least one of the intermediate layers;
generating, using the intercepted activations and reference activations, a first metric based on a Fréchet distance;
generating, using the intercepted activations and the reference activations, a second metric based on a class-conditioned Fréchet distance; and
initiating a remediation action to protect the AI model from a privacy attack when the first metric is above a first threshold and the second metric is above a second threshold.
27. A computer-implemented method comprising:
inputting each of a plurality of queries into an artificial intelligence (AI) model comprising a plurality of layers including an input layer, an output layer, and at least one intermediate layer between the input layer and the output layer, each intermediate layer, during inference, outputting a plurality of activations;
intercepting, for each query, activations from at least one of the intermediate layers;
determining whether a distribution of the intercepted activations across the queries indicates that the queries seek to cause the AI model to behave in an undesired manner by conducting a distance-based similarity analysis between the intercepted activations and reference activations, the distance-based similarity analysis utilizing a combination of a Fréchet distance and a class-conditioned Fréchet distance;
identifying, based on the determination, a data point in a training data set used by the AI model subject to a privacy attack; and
providing data characterizing the identified data point to a consuming application or process.
28. A computer-implemented method comprising:
intercepting each of a plurality of queries of an artificial intelligence (AI) model;
inputting the intercepted queries into a surrogate model corresponding to the AI model, the surrogate model comprising a plurality of layers including an input layer, an output layer, and at least one intermediate layer between the input layer and the output layer, each intermediate layer, during inference, outputting a plurality of surrogate activations, the surrogate model being a more compact representation of the AI model;
intercepting, for each query, surrogate activations from at least one of the intermediate layers;
determining whether a distribution of the intercepted surrogate activations across the queries indicates that the queries seek to cause the AI model to behave in an undesired manner by conducting a distance-based similarity analysis between the intercepted surrogate activations and reference surrogate activations; and
providing data characterizing the determining to a consuming application or process.
29. The method of claim 28 further comprising:
initiating, by the consuming application or process, at least one remediation action to prevent the AI model from behaving in an undesired manner.