Dria

Summary

The scoring mechanism in the finalizeValidation function of LLMOracleCoordinator.sol is vulnerable to manipulation by malicious validators who can skew the final scores and unfairly obtain validator rewards. This occurs because the function calculates rewards based on scores within a range defined by the mean and standard deviation of all scores, without robust protections against coordinated low scores. Malicious validators can take advantage of this flaw to drive down the mean, ensuring only their low scores fall within the valid range for rewards, while legitimate, higher scores are excluded.

Vulnerability Details

tep-by-Step Attack Scenario

Prerequisites
Consider a task with the following parameters:

numValidations = 5 (5 validators required)
3 malicious validators
2 honest validators
Maximum score: 100
Attack Execution

Initial Setup

Task ID: 1
Maximum score: 100
Number of validators: 5
Validator fee: 1000 tokens
Honest Validators Submit First

Honest Validator 1 submits: 75
Honest Validator 2 submits: 80
Malicious Validators Coordinate

Malicious Validator 1 submits: 20
Malicious Validator 2 submits: 25
Malicious Validator 3 submits: 30
Score Calculation

All Scores: [75, 80, 20, 25, 30]
Mean = (75 + 80 + 20 + 25 + 30) / 5 = 46
Standard Deviation ≈ 28
Valid Range (mean ± stddev): 18 to 74
Result

Check which scores fall within the valid range (18 to 74):
20 - Within range ✅ (Malicious gets paid)
25 - Within range ✅ (Malicious gets paid)
30 - Within range ✅ (Malicious gets paid)
75 - Outside range ❌ (Honest doesn’t get paid)
80 - Outside range ❌ (Honest doesn’t get paid)
Final Score Calculation

inner_score = (20 + 25 + 30) / 3 = 25
This results in a low score even though higher scores were submitted.

Impact

Honest validators are deprived of their rightful rewards (1000 tokens each lost).
The manipulated score does not accurately represent high-quality evaluations.
Honest validators are discouraged from participating due to unfair losses.

Tools Used

Manual Review

Recommendations

Use a reputation-weighted scoring mechanism, where validators’ reputation influences their impact on the final score.