Deep learning models that predict drug--target interactions (DTI) can speed up drug discovery, but the candidate drugs they score are valuable intellectual property that their owners cannot reveal. Secure multiparty computation (MPC) lets two servers compute a prediction on a secret-shared drug, without either server learning the drug. However, secure inference of deep models is slow. Non-linear operations require communication between the servers, and MPC arithmetic does not map naturally to GPUs. In this paper, we present our solution to Track~III of the iDASH 2026 privacy and security competition, which asked for fast two-party secure DTI inference on GPUs. Our solution is built on the co-design of MPC-friendly models and MPC protocols. On the model side, we design SlimDTA, a DTI model that reduces the operations on the secret drug, and in particular the non-linear operations that require communication. On the protocol side, we build a GPU framework that evaluates the linear layers with additive secret sharing and the non-linear layers with function secret sharing, including a custom CUDA kernel for the comparisons. Our solution won first place in the competition, at under 10,ms per prediction on average, and its secure predictions match the plaintext model to within 0.02%.
Figures & tables
Figure 1. Secure DTI inference as a service. The drug owner splits its secret drug into two shares and sends one to each server. The protein and the model are public, and both servers know them. The servers run the secure protocol on their GPUs and exchange only masked values, which look uniformly random. Each server returns its share of the prediction, and only the drug owner can add the two shares to obtain y^ . A trusted dealer prepares masks, keys, and triples offline.
Symbol
Meaning
P0,P1
the two servers
Z2ℓ
the ring of integers modulo 2ℓ ; we use ℓ=64
f
number of fractional bits in fixed-point encoding; we use f=16
[[x]]=(x0,x1)
additive secret sharing of x ; server Pi holds xi
MSB(x)
most significant bit of x ; equals 1 if and only if x is negative
r , x^=x+r
a random mask and a masked value
Table 1. Notation.
Figure 2. Additive secret sharing in Z256 . The secret x=5 is split into two shares that look random on their own. To compute the linear function y=3x+2 , each server applies it to its own share (only P0 adds the constant). The two results add up to the correct answer, without any communication.
Figure 3. A secure ReLU with function secret sharing. Red steps communicate and green steps are local. The expensive part of a comparison, step 2, is a local computation on the GPU. The servers only ever see the random-looking value x^ .
Figure 4. Evaluating a distributed comparison function on 3-bit inputs. The keys define a binary tree whose path to the secret threshold α (orange, dashed) separates the leaves with z<α from the rest. A server evaluates its key at a public input z by walking a single root-to-leaf path (blue). The two servers’ outputs add up to g(z)=1[z<α] , here 1 , yet neither server learns α . In our solution the tree has 63 levels.
Protocol
Description
Rounds
Bits sent
Reference
Basic primitives
[[x]]+[[y]] , c+[[x]] , c[[x]]
addition, and operations with a public constant
0
0
Eqs. ( 4 )–( 6 )
π\textscTrunc
division by 2f after a product
0
0
Mohassel and Zhang (2017)
π\textscOpen
reveal x to both servers
1
ℓ
—
π\textscMul
product [[x⋅y]] with a Beaver triple
1
2ℓ
Beaver (1991)
π\textscSign
XOR shares of 1[x≥0] , by FSS
1
ℓ
Boyle et al. (2021)
Table 2. The protocols used in our solution and their online cost, in three levels. Bits are those sent by each server per output element, with ℓ=64 . W , b , and c are public, and L is the number of values in a max-pooling window.
Figure 5. SlimDTA, with shapes (channels × positions) for Davis. Block heights are proportional to channel counts (fully connected layers are not to scale), and “max” is a global maximum over positions. Orange marks everything that depends on the drug, the only secret. Blue marks the protein tower, which depends only on public information. The protein tower performs 98% of the plaintext multiply-adds, but needs no secure computation.
DeepDTA ( Öztürk et al., 2018 )
DeepDTAGen (affinity)
SlimDTA
Drug input
SMILES
molecular graph
SMILES
Drug encoder
3 conv (32, 64, 96)
3 graph conv
3 conv (16, 32, 48)
Protein encoder
3 conv (32, 64, 96)
gated conv, flattened into
3 dilated conv (32, 64, 96)
a 1.3M-parameter linear layer
Head
3 FC layers
4 FC layers
2 FC layers
Parameters
1.9M
3.58M
0.80M
Table 3. SlimDTA compared with DeepDTA and DeepDTAGen. Balanced accuracy is the mean over 7 affinity thresholds on the 5,010 Davis and 19,653 KIBA test pairs. ∗ Measured by running the released DeepDTAGen checkpoint on the same split. The SlimDTA numbers are for the checkpoint we submitted, our best training run. Over training seeds, its Davis balanced accuracy averages about 0.82.
Figure 6. Dilated convolution with k=3 taps. The same three weights (dark cells) are spaced d positions apart, so a larger d covers a wider window, (k−1)⋅d+1 positions, with no extra weights. In SlimDTA, the protein tower is public, so the wider view adds no secure computation.
Figure 7. The secure evaluation of SlimDTA. Orange marks values that depend on the drug, and blue marks public values. Layers with a green border are linear with public weights and run locally on each server’s GPU. Layers with a red border are comparisons evaluated with FSS, and are the only layers that communicate. The protein tower (blue) runs in plaintext and joins at the first fully connected layer through the split of Equation ( 20 ). The rounds do not depend on the batch size.
Figure 8. Computing a product modulo 264 with limbs. Each 64-bit value is split into four 16-bit limbs (left). Of the 16 limb products, only the 10 with i+j<4 contribute modulo 264 (right). Each kept product is an exact floating-point matrix product on the GPU.
Figure 9. Two ways to evaluate a DCF key on the GPU. A tree walk written with PyTorch tensor operations launches several kernels per level and writes the intermediate seeds of every comparison to global memory, only to read them back at the next level. Our fused kernel walks all 63 levels in one launch, with the seed and the running value in registers. In both cases, each GPU thread handles one comparison, and all comparisons of a layer run in parallel.
Team
BindingDB (1,000 samples)
BindingDB (10,000 samples)
Davis (1,000 samples)
PPMLHuskies (ours)
9.2 s
91.5 s
8.1 s
Team A
14.4 s
106.1 s
9.5 s
Team B
14.0 s
164.6 s
12.7 s
Team C
2,486.0 s
25,463.5 s
2,518.9 s
Ours, per sample
9.2 ms
9.2 ms
8.1 ms
Ours, speedup over the next fastest
1.52 ×
1.16 ×
1.17 ×
Table 4. Secure inference time in the iDASH 2026 Track III evaluation, as reported by the organizers. Two H100 workstations, a network of more than 1 Gbps, and batch size 64.
Figure 10. Online time per sample for real two-party inference over a 1 Gbps link with 0.25 ms delay, on the Davis model (log scales). Our solution approaches the communication time, shown dashed, which no amount of GPU computation can reduce.
Online time per sample (ms)
Ours, split (ms)
Batch
FSS + CUDA (ours)
FSS + PyTorch
Bit decomposition
Communication
Computation
16
67.8
293.4
386.3
10.6
56.9
64
20.9
72.8
252.0
8.9
12.0
256
11.5
33.7
223.4
8.6
3.0
512
10.0
32.1
219.0
8.5
1.5
1,024
9.5
34.7
out of memory
8.6
0.9
Table 5. Online time per sample on Davis, real two-party inference over a 1 Gbps link. The right two columns split our time into communication and computation.
Figure 11. Where our online time goes. Computation (green) shrinks as the batch grows. Communication (red) stays at about 8.6 ms per sample, the time for both servers to send 0.505 MB per sample each over the shared 1 Gbps link.
Secret inputs
Comparisons by
Comparisons per sample
Rounds
MB per sample
drug, protein, weights
bit decomposition
≈435,000
1,727
481
drug, protein, weights
FSS
≈435,000
90
41.5
drug only
bit decomposition
12,624
1,549
13.2
drug only
FSS (ours)
12,624
37
0.505
Table 6. Communication of secure SlimDTA inference on Davis under different settings. Rounds are per batch and do not depend on the batch size. Megabytes are sent by each server, per sample.
Layer
Output
Secret multiply-adds
Comparisons
Rounds
Embedding OE
85×128
707,200
0
0
Conv 1 (kernel 4) and ReLU
16×82
671,744
1,312
3
Conv 2 (kernel 4) and ReLU
32×79
161,792
2,528
3
Conv 3 (kernel 4) and ReLU
48×76
466,944
3,648
3
Global max pool
48
0
3,600
21
FC 1, drug part ( 48→1024 ), and ReLU
1024
49,152
1,024
3
Table 7. Secure cost of SlimDTA per sample on Davis. All multiply-adds are local. Rounds are per batch.
Department of Economics and Management, University of Pavia, Pavia, Italy · Department of Computer Science and Engineering, Seoul National University, Gwanak-gu 08826, Korea.