Open Access

Table 3.

Comparison of various PPML protocols

PPML Capability
Threat Model
Techniques Neural
Inference Training Semi-honest Malicious Networks
2PC SecureML [24] HE/GC/ASS From [24]
2PC MiniONN [277] HE/GC/ASS From [24, 277]
2PC GAZELLE [20] HE/GC/ASS From [24, 277]
2PC EzPC [278] GC/ASS From [24, 277]
2PC XONN [29] GC/ASS VGG-16 [279]
2PC QUOTIENT [18] OT/GC/ASS From [18]
2PC MP2ML [280] HE/GC/ASS CryptoNets [281]
2PC CrypTFlow2 [28] HE/OT/ASS DenseNet-121 [282]
2PC Delphi [22] HE/GC/ASS VGG-16 [279]
2PC QuantizedNN [283] Abort HE/OT/ASS MobileNets [284]
2PC SIRNN [27] OT/ASS Heads [285]

3PC Chameleon [286] GC/ASS AlexNet [287]
3PC ABY3 [23] GC/ASS From [24, 277]
3PC ASTRA [288] Abort ASS/RSS From [24]
3PC SecureNN [289] ASS From [24, 277]
3PC BLAZE [26] Fairness ASS/RSS From [24]
3PC QuantizedNN [283] Abort RSS MobileNets [284]
3PC CrypTFlow [21] ASS DenseNet-121 [282]
3PC SWIFT [290] GOD ASS/RSS VGG-16 [279]
3PC CryptGPU [31] RSS ResNet-152 [291]
3PC Falcon [292] Abort RSS VGG-16 [279]

4PC FLASH [293] GOD ASS/RSS From [24]
4PC SWIFT [290] GOD ASS/RSS VGG-16 [279]
4PC Trident [19] Fairness GC/ASS/RSS From [24]
4PC Tetrad [294] GOD GC/ASS/RSS VGG-16 [279]

Note: All protocols for secure three-party/four-party computation (i.e., 3PC/4PC) tolerate one corruption, and thus belong to the honest-majority setting. For malicious adversaries, “Abort”, “Fairness”, and “GOD” denote the PPML protocols that achieve security with abort, fairness, and guaranteed output delivery, respectively. For the underlying LSSS, we use “ASS” and “RSS” to denote the additive secret sharing and the replicated secret sharing, respectively. If a PPML protocol supports multiple neural-network architectures, we only describe the one with largest parameters for private ML inference.

Current usage metrics show cumulative count of Article Views (full-text article views including HTML views, PDF and ePub downloads, according to the available data) and Abstracts Views on Vision4Press platform.

Data correspond to usage on the plateform after 2015. The current usage metrics is available 48-96 hours after online publication and is updated daily on week days.

Initial download of the metrics may take a while.