AuthGuard-7702
Screening delegate bytecode before account authorization
M M Imran, Anik Tahabilder, Changxin Bai, Qiang Zhu, Shiyong Lu
8th IEEE International Conference on Trust, Privacy and Security in Intelligent Systems, and Applications (TPS), 2026. Accepted.
Problem
An Ethereum account can delegate execution to smart-contract code. Before a user signs that authorization, the proposed code may have little reputation or transaction history. An attacker can also modify the bytecode in response to a screening service.
Approach
AuthGuard-Seq splits the opcode sequence into chunks, encodes each chunk, and learns which chunks matter most for the decision. Contracts from the same family never appear in both training and test data. The attacker combines bytecode transformations, and transformed programs are executed in a real EVM to check that they still behave like the original.
Research contribution
On ordinary test data, AuthGuard-Seq and a simple logistic-regression screener are not statistically different. Against an adaptive attacker whose changes are checked to preserve behavior, AuthGuard-Seq holds up much better. The paper argues that screeners should be compared this way, not only on clean data.
Authorship
First author, with Anik Tahabilder, Changxin Bai, Qiang Zhu, and Shiyong Lu.
Evaluation evidence
On 2,190 EIP-7702 delegates, the paper reports paired attack success of 0.167 for AuthGuard-Seq versus 0.537 for a 15-feature logistic-regression screener under the primary adversary. Robust recall is 0.679 versus 0.378. The primary attack space is supported by 5,196 transformed executions from 120 bytecode families.
Scope & limitations
The results hold for the attack target, transformation set, query budget, and execution checks defined in the paper. They do not show that AuthGuard-Seq detects malicious contracts in general, or that transformed programs are fully equivalent to the originals. On clean data, AuthGuard-Seq is not statistically better than logistic regression. The later scan of real delegation traffic was unlabeled, so it does not measure detection accuracy in the field.
Reproduction
The repository holds several rounds of evaluation code. Use the paper and the revision notes in the repository to pick the right experiment; there is not yet a single tagged release that matches the accepted paper.
Citation
M M Imran, Anik Tahabilder, Changxin Bai, Qiang Zhu, Shiyong Lu. “Pre-Authorization Screening of EIP-7702 Account Delegation in Ethereum Blockchain under an Execution-Audited Adaptive Adversary.” 8th IEEE International Conference on Trust, Privacy and Security in Intelligent Systems, and Applications (TPS), 2026. Accepted.
BibTeX
@inproceedings{authguard77022026,
title = {Pre-Authorization Screening of EIP-7702 Account Delegation in Ethereum Blockchain under an Execution-Audited Adaptive Adversary},
author = {M M Imran and Anik Tahabilder and Changxin Bai and Qiang Zhu and Shiyong Lu},
booktitle = {8th IEEE International Conference on Trust, Privacy and Security in Intelligent Systems, and Applications (TPS)},
year = {2026},
note = {Accepted},
}Contact
If you'd like to talk about a research collaboration, a faculty position, or an industry research role in blockchain security, email me.
mmimran@wayne.edu