Table 1.
Representative secure Transformer models inference frameworks. MM is for matrix multiplication, L.N. indicates LayerNorm, and E.M. stands for Embedding. • indicates the used techniques and ° is for not used. ✔ indicates supporting a feature and ✘ is for not supported ones. M-Mod is for multi-modality. Compared to two-party computation (2PC) frameworks, three-party computation (3PC) ones are usually faster at the cost of additional assumptions of non-colluding servers
| Framework | Techniques |
Transformer Functions |
Re-Train | M-Mod. | Code | ||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| HE | SS | OT | FSS | MM | GELU | Softmax | L.N. | E.M. | |||||
| 2PC | IRON [14] | • | • | • | ° | ✔ | ✔ | ✔ | ✔ | ✘ | Not Req. | ✘ | ✔ |
| BumbleBee [12] | • | • | • | ° | ✔ | ✔ | ✔ | ✔ | ✔ | Not Req. | ✘ | ✔ | |
| CipherGPT [15] | • | • | • | ° | ✔ | ✔ | ✔ | ✔ | ✔ | Not Req. | ✘ | ✘ | |
| BOLT [16] | • | • | • | ° | ✔ | ✔ | ✔ | ✔ | ✔ | Req. | ✘ | ✔ | |
| NEXUS [23] | • | ° | ° | ° | ✔ | ✔ | ✔ | ✔ | ✘ | Not Req. | ✘ | ✔ | |
| 3PC | SIGMA [13]† | ° | ° | ° | • | ✔ | ✔ | ✔ | ✔ | ✘ | Not Req. | ✘ | ✔ |
| Privformer [17] | ° | • | ° | ° | ✔ | ✘ | ✘ | ✔ | ✔ | Req. | ✘ | ✘ | |
| Ditto [20] | ° | • | ° | ° | ✔ | ✔ | ✔ | ✔ | ✔ | Req. | ✘ | ✔ | |
| CRYPTEN [24]† | ° | • | ° | ° | ✔ | ✘ | ✔ | ✘ | ✘ | Req. | ✘ | ✔ | |
| MPCFORMER [11]† | ° | • | ° | ° | ✔ | ✘ | ✘ | ✘ | ✘ | Req. | ✘ | ✔ | |
| PUMA (Ours) | ° | • | ° | ° | ✔ | ✔ | ✔ | ✔ | ✔ | Not Req. | ✔ | ✔ | |
† They work in dealer-based 3PC, where the dealer generates the correlated randomness for two computing parties.
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.