Privacy-Preserving Machine Learning

Proposed a fully decentralized protocol with differential privacy guarantees in which no central trusted entity exists and all clients are honest-but- curious.
Proposed differentially private collaborative machine learning algorithms for classification and regression tasks.
Run experiments to show that the accuracy of the proposed algorithms trained with practical datasets closely follows the result of centralized training performed by a single trusted entity with a negligible loss.

System Design and Evaluation for Power-Domain NOMA

Proposed a novel triangular QAM based constellation design for two-user uplink NOMA systems based on a minimum distance criterion with both users deploy constellations of arbitrary even-power-of-two points.
Modified the proposed constellation structure of downlink NOMA systems with channel coding evaluations. When the power ratio between the two users in the system is approaching unity, the performance gain for the strong user becomes larger than 4dB compared to the second-best design.

Designed a scheme to replace the widely-used one-hot encoded labels with codewords generated by conventional error-correcting BCH codes.
Utilized frameworks such as PyTorch and Keras via a convolutional neural network structure, ResNet, to verify that the BCH encoded label can enhance the accuracy in classification problems, e.g., age estimation problems.

Designed an effective algorithm for exact error probability computation under binary erasure channel (BEC) of each bit-channel in arbitrarily given kernels.
Constructed simulations and indicated that a few 4*4 kernels achieve better performances compared to the conventional Kronecker-extended 4*4 kernel in both BEC and AWGN channel.
Identified that with a better performance in the selected 4*4 kernel, their Kronecker-extended structures with a size of 16*16, however, result in worse performance compared to the conventional kernel in both BEC and AWGN channel.

Page updated

Report abuse