Deep learning unlocks applications with societal impacts, e.g., detecting child exploitation imagery and genomic analysis of rare diseases. Deployment, however, needs compliance with stringent privacy regulations. Training algorithms that preserve the privacy of training data are in pressing need. Purely cryptographic approaches can protect privacy, but they are still costly, even when they rely on two or more non-colluding servers. Seemingly-"trivial" operations in plaintext quickly become prohibitively inefficient when a series of them are "crypto-processed," e.g., (dynamic) quantization for ensuring the intermediate values would not overflow. Slalom, recently proposed by Tramer and Boneh, is the first solution that leverages both GPU (for efficient batch computation) and a trusted execution environment (TEE) (for minimizing the use of cryptography). Roughly, it works by a lot of pre-computation over known and fixed weights, and hence it only supports private inference. Five related problems for private training are left unaddressed. Goten, our privacy-preserving training and prediction framework, tackles all five problems simultaneously via our careful design over the "mismatched" cryptographic and GPU data types (due to the tension between precision and efficiency) and our round-optimal GPU-outsourcing protocol (hence minimizing the communication cost between servers). It 1) stochastically trains a low-bitwidth yet accurate model, 2) supports dynamic quantization (a challenge left by Slalom), 3) minimizes the memory-swapping overhead of the memory-limited TEE and its communication with GPU, 4) crypto-protects the (dynamic) model weight from untrusted GPU, and 5) outperforms a pure-TEE system, even without pre-computation (needed by Slalom). As a baseline, we build CaffeScone that secures Caffe using TEE but not GPU; Goten shows a 6.84x speed-up of the whole VGG-11. Goten also outperforms Falcon proposed by Wagh et al., the latest secure multi-server cryptographic solution, by 132.64x using VGG-11. Lastly, we demonstrate Goten's efficacy in training models for breast cancer diagnosis over sensitive images.

No abstract

Privacy-preserving deep neural network (DNN) inference remains an intriguing problem even after the rapid developments of different communities. One challenge is that cryptographic techniques such as homomorphic encryption (HE) do not natively support non-linear computations (e.g., sigmoid). A recent work, BAYHENN (Xie et al., IJCAI'19), considers HE over the Bayesian neural network (BNN). The novelty lies in "meta-prediction" over a few noisy DNNs. The claim was that the clients can get intermediate outputs (to apply non-linear function) but are still prevented from learning the exact model parameters, which was justified via the widely-used learning-with-error (LWE) assumption (with Gaussian noises as the error). This paper refutes the security claim of BAYHENN via both theoretical and empirical analyses. We formally define a security game with different oracle queries capturing two realistic threat models. Our attack assuming a semi-honest adversary reveals all the parameters of single-layer BAYHENN, which generalizes to recovering the whole model that is "as good as" the BNN approximation of the original DNN, either under the malicious adversary model or with an increased number of oracle queries. This shows the need for rigorous security analysis ("the noise introduced by BNN can obfuscate the model" fails -- it is beyond what LWE guarantees) and calls for the collaboration between cryptographers and machine-learning experts to devise practical yet provably-secure solutions.

No abstract

No abstract