Differentially private training offers a protection which is usually interpreted as a guarantee against membership inference attacks. By proxy, this guarantee extends to other threats like reconstruction attacks attempting to extract complete training examples. Recent works provide evidence that if one does not need to protect against membership attacks but instead only wants to protect against a training data reconstruction, then utility of private models can be improved because less noise is required to protect against these more ambitious attacks. We investigate this question further in the context of DP-SGD, a standard algorithm for private deep learning, and provide an upper bound on the success of any reconstruction attack against DP-SGD together with an attack that empirically matches the predictions of our bound. Together, these two results open the door to fine-grained investigations on how to set the privacy parameters of DP-SGD in practice to protect against reconstruction attacks. Finally, we use our methods to demonstrate that different settings of the DP-SGD parameters leading to same DP guarantees can results in significantly different success rates for reconstruction, indicating that the DP guarantee alone might not be a good proxy for controlling the protection against reconstruction attacks.

Differentially private training offers a protection which is usually interpreted as a guarantee against membership inference attacks. By proxy, this guarantee extends to other threats like reconstruction attacks attempting to extract complete training examples. Recent works provide evidence that if one does not need to protect against membership attacks but instead only wants to protect against a training data reconstruction, then utility of private models can be improved because less noise is required to protect against these more ambitious attacks. We investigate this question further in the context of DP-SGD, a standard algorithm for private deep learning, and provide an upper bound on the success of any reconstruction attack against DP-SGD together with an attack that empirically matches the predictions of our bound. Together, these two results open the door to fine-grained investigations on how to set the privacy parameters of DP-SGD in practice to protect against reconstruction attacks.

Being able to reconstruct training data from the parameters of a neural network is a major privacy concern. Previous works have shown that reconstructing training data, under certain circumstances, is possible. In this work, we analyse such reconstructions empirically and propose a new formulation of the reconstruction as a solution to a bilevel optimisation problem. We demonstrate that our formulation as well as previous approaches highly depend on the initialisation of the training images $x$ to reconstruct. In particular, we show that a random initialisation of $x$ can lead to reconstructions that resemble valid training samples while not being part of the actual training dataset. Thus, our experiments on affine and one-hidden layer networks suggest that when reconstructing natural images, yet an adversary cannot identify whether reconstructed images have indeed been part of the set of training samples.