Summary of Differential Privacy for Wireless Federated Learning (2025)
Published:
Paper List
Differential Privacy for Wireless Federated Learning
2. DP-FL-Power, Globecom, 2020
3. DP-Privacy-for-Free, JSAC, 2020
8. DP-OTA-FL-Device-Sampling, Globecom, 2023
9. DP-OTA-FL-Device-Sampling+, TWC, 2024
10. DP-OTA-FL-Random, TWC, 2024
11. DP-OTA-FL-MIMO, Globecom, 2023
1. WFL-LDP, ISIT, 2020
Seif M, Tandon R, Li M. Wireless federated learning with local differential privacy[C]//2020 IEEE International Symposium on Information Theory (ISIT). IEEE, 2020: 2604-2609.
- sample-level local differential privacy
- time-invariant channel coefficients
- gradient descent
- privacy analysis:
- bounded sample-wise gradient
- advanced composition
convergence analysis:
smoothness
strongly convex
2. DP-FL-Power, Globecom, 2020
Koda Y, Yamamoto K, Nishio T, et al. Differentially private aircomp federated learning with power adaptation harnessing receiver noise[C]//GLOBECOM 2020-2020 IEEE Global Communications Conference. IEEE, 2020: 1-6.
- sample-level differential privacy
- control the power-scaling factor to satisfy DP constraint
- privacy analysis (per-iteration):
- standard Gaussian mechanism analysis
- SNR-privacy-level trade-off
3. DP-Privacy-for-Free, JSAC, 2020
[1] Liu D, Simeone O. Privacy for free: Wireless federated learning via uncoded transmission with adaptive power control[J]. IEEE Journal on Selected Areas in Communications, 2020, 39(1): 170-185.
[2] Liu D, Sonee A, Simeone O, et al. 17 Differentially Private Wireless Federated Learning[J]. Machine Learning and Wireless Communications, 2022: 486.
- sample-level differential privacy
- OMA and NOMA, block flat-fading
- gradient descent
- privacy analysis:
- bounded sample-wise gradient
- advanced composition
- convergence analysis:
- smoothness
- PL
4. FedSGD-DP/LDP, JSAC, 2021
Mohamed M S E, Chang W T, Tandon R. Privacy amplification for federated learning via user sampling and wireless aggregation[J]. IEEE Journal on Selected Areas in Communications, 2021, 39(12): 3821-3835.
- sample-level (local) differential privacy
- analyze the effect of user sampling
- privacy analysis:
- bounded sample-wise gradient
- privacy amplification by subsampling-based analysis (such as Hoeffding’s Inequality)
- advanced composition
- convergence analysis:
- smoothness
- strongly convex
5. PWFL-OAC, ICASSP, 2021
Hasırcıoğlu B, Gündüz D. Private wireless federated learning with anonymous over-the-air computation[C]//ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021: 5195-5199.
- sample-level differential privacy
- analyze the effect of user sampling and SGD sampling
- privacy analysis:
- sampled Gaussian mechanism analysis
- RDP composition
6. S-DPOTAFL, ICC, 2023
Yan N, Wang K, Pan C, et al. Device scheduling for over-the-air federated learning with differential privacy[C]//ICC 2023-IEEE International Conference on Communications. IEEE, 2023: 51-56.
- sample-level differential privacy
- gradient descent
- time-invariant channel coefficient
- privacy analysis (per client, per iteration):
- bounded sample-wise gradient
- Gaussian mechanism analysis
- convergence:
- smoothness
- PL
- optimization
- optimize the scheduled devices set and alignment coefficient
- analyze the condition that S-DPOTAFL performs better than the DP-OTA-FL without considering device scheduling (NoS-DPOTAFL)
7. S-DPOTAFL+, IOTJ, 2024
Yan N, Wang K, Zhi K, et al. Device Scheduling for Secure Aggregation in Wireless Federated Learning[J]. IEEE Internet of Things Journal, 2024.
- sample-level differential privacy
- some set of the devices scheduled to send artificial noise
- privacy analysis (per client, per iteration):
- bounded sample-wise gradient
- Gaussian mechanism analysis
- convergence:
- smoothness
- strongly convex
- device scheduling policies
- sufficient channel noise for all device participation
- sufficient channel noise for partial device participation
- select those devices as participants to engage in the training, and other devices will be absent in this round
- select some devices as participants and others are selected as helpers to send artificial noise
- insufficient channel noise for any device participation
- some devices need to be selected as helpers to send artificial noise
- optimization
- optimize device scheduling for the insufficient channel noise case
- integer nonlinear fractional optimization problem
- branch-and-bound (BnB)-based algorithm
8. DP-OTA-FL-Device-Sampling, Globecom, 2023
Hu Z, Yan J, Zhang Y J A. Towards differentially private over-the-air federated learning via device sampling[C]//GLOBECOM 2023-2023 IEEE Global Communications Conference. IEEE, 2023: 5292-5298.
- sample-level differential privacy
- time-invariant channel coefficient
- privacy analysis:
- bounded sample-wise gradient
- moment accountant
- convergence:
- smoothness
- convex
- optimization
- optimize the power allocation and the number of device sampling set
- derive the optimal transceiver design
- derive the optimal number of sampling trials
9. DP-OTA-FL-Device-Sampling+, TWC, 2024
Hu Z, Yan J, Zhang Y J A. Communication-learning co-design for differentially private over-the-air federated learning with device sampling[J]. IEEE Transactions on Wireless Communications, 2024.
- sample-level differential privacy
- time-invariant channel coefficient
- inactive devices can choose to transmit artificial noise
- privacy analysis:
- bounded sample-wise gradient
- moment accountant
- convergence:
- smoothness
- convex
- device scheduling policies
- optimize the power allocation and the number of device sampling set
- derive the optimal transceiver design
- derive the optimal number of sampling trials
- further discussion:
- non-convex convergence
- multiple local iterations
10. DP-OTA-FL-Random, TWC, 2024
Park S, Choi W. On the differential privacy in federated learning based on over-the-air computation[J]. IEEE Transactions on Wireless Communications, 2023, 23(5): 4269-4283.
- sample-level differential privacy
- quasi-static fading channel
- consider inherent randomness of the local gradient (CLT -> Gaussian mechanism)
- privacy analysis:
- bounded sample-wise gradient
- advanced composition
- convergence:
- smoothness
- need parameter estimation (such as the gradient norm and the covariance of random gradient)
11. DP-OTA-FL-MIMO, Globecom, 2023
Liu H, Yan J, Zhang Y J A. On the privacy leakage of over-the-air federated learning over MIMO fading channels[C]//GLOBECOM 2023-2023 IEEE Global Communications Conference. IEEE, 2023: 5274-5279.
- sample-level differential privacy
- time-invariant channel coefficient
- with a multiple-antenna server
- FL model aggregation with a multiple-antenna server amplifies privacy leakage
- privacy analysis:
- bounded sample-wise gradient
- moment accountant
- private information extraction optimization through receiver beamformer
- free DP can not be attained without the addition of extra artificial noise for the MIMO system
12. DP-OTA-FL-MIMO, TWC, 2024
Liu H, Yan J, Zhang Y J A. Differentially private over-the-air federated learning over MIMO fading channels[J]. IEEE Transactions on Wireless Communications, 2024, 23(8): 8232-8247.
- sample-level differential privacy
- time-invariant channel coefficient
- with a multiple-antenna server
- FL model aggregation with a multiple-antenna server amplifies privacy leakage
- privacy analysis:
- bounded sample-wise gradient
- moment accountant
- private information extraction optimization through receiver beamformer
free DP can not be attained without the addition of extra artificial noise for the MIMO system
- further discussion:
- convergence
- strongly convex
- smoothness
- privacy-convergence optimization
- convergence
13. Others
- Non-coherent
Seif M, Şahin A, Poor H V, et al. On Differential Privacy for Wireless Federated Learning with Non-coherent Aggregation[C]//GLOBECOM 2023-2023 IEEE Global Communications Conference. IEEE, 2023: 213-218.