notesum.ai
Published at November 9Smart-LLaMA: Two-Stage Post-Training of Large Language Models for Smart Contract Vulnerability Detection and Explanation
cs.CR
cs.AI
cs.SE
Released Date: November 9, 2024
Authors: Lei Yu1, Shiqi Chen, Hang Yuan1, Peng Wang1, Zhirong Huang1, Jingyuan Zhang1, Chenjie Shen1, Fengjun Zhang2, Li Yang2, Jiajia Ma2
Aff.: 1Institute of Software, Chinese Academy of Sciences, Beijing, China; 2University of Chinese Academy of Sciences, Beijing, China

| Methods | Reentrancy | Timestamp Dependency | Overflow/Underflow | Delegatecall | ||||||||||||
| A(%) | P(%) | R(%) | F1(%) | A(%) | P(%) | R(%) | F1(%) | A(%) | P(%) | R(%) | F1(%) | A(%) | P(%) | R(%) | F1(%) | |
| Mythril | 54.21 | 33.75 | 73.97 | 46.35 | 43.27 | 44.63 | 44.13 | 44.38 | 37.45 | 25.30 | 46.67 | 32.81 | 60.71 | 42.99 | 74.19 | 54.44 |
| Osiris | 30.77 | 26.80 | 91.78 | 41.49 | 51.03 | 52.50 | 36.63 | 43.15 | 61.90 | 45.33 | 75.56 | 56.67 | – | – | – | – |
| Oyente | 68.50 | 42.35 | 49.32 | 45.57 | 53.98 | 66.67 | 18.60 | 29.09 | 72.53 | 60.87 | 46.67 | 52.83 | 64.71 | 40.00 | 31.58 | 35.29 |
| Slither | 37.73 | 16.08 | 31.51 | 21.30 | 57.02 | 56.50 | 70.39 | 62.69 | 50.91 | 32.28 | 45.56 | 37.79 | 52.04 | 39.04 | 91.94 | 54.81 |
| Smartcheck | 43.22 | 30.10 | 84.93 | 44.44 | 51.00 | 57.14 | 17.88 | 27.23 | 52.00 | 31.25 | 38.89 | 34.65 | 54.08 | 32.93 | 43.55 | 37.50 |
| GCN | 73.21 | 74.47 | 73.18 | 73.82 | 75.91 | 74.93 | 77.55 | 76.22 | 67.53 | 69.52 | 70.93 | 70.22 | 65.76 | 69.01 | 69.74 | 69.37 |
| TMP | 76.45 | 76.04 | 75.30 | 75.67 | 78.84 | 78.68 | 76.09 | 77.36 | 70.85 | 70.26 | 69.47 | 69.86 | 69.11 | 68.18 | 70.37 | 69.26 |
| AME | 81.06 | 79.62 | 78.45 | 79.03 | 82.25 | 81.42 | 80.26 | 80.84 | 73.24 | 71.36 | 71.59 | 71.47 | 72.85 | 70.25 | 69.40 | 69.82 |
| SMS | 83.85 | 79.46 | 77.48 | 78.46 | 89.77 | 89.15 | 91.09 | 90.11 | 79.36 | 78.14 | 72.98 | 75.47 | 78.82 | 76.97 | 73.69 | 75.29 |
| DMT | 89.42 | 83.62 | 81.06 | 82.32 | 94.58 | 93.60 | 96.39 | 94.97 | 85.64 | 85.44 | 74.32 | 79.49 | 82.76 | 84.61 | 77.93 | 81.13 |
| Peculiar | 58.72 | 35.23 | 48.44 | 40.84 | 69.08 | 77.31 | 71.12 | 74.18 | 75.91 | 64.29 | 60.00 | 62.07 | 88.46 | 77.36 | 82.00 | 79.61 |
| PSCVFinder | 58.26 | 35.16 | 50.00 | 41.32 | 40.10 | 52.00 | 50.39 | 51.18 | 51.09 | 33.58 | 50.00 | 40.18 | 89.56 | 79.25 | 84.00 | 81.57 |
| LLaMA-3.1-8B | 36.24 | 29.73 | 85.94 | 44.18 | 57.00 | 67.24 | 60.47 | 63.67 | 55.11 | 38.93 | 64.44 | 48.54 | 63.74 | 38.89 | 56.00 | 45.90 |
| Qwen2-7B | 30.28 | 29.63 | 100.00 | 45.71 | 67.63 | 71.83 | 79.07 | 75.28 | 66.42 | 47.62 | 22.22 | 30.30 | 65.93 | 44.23 | 92.00 | 59.74 |
| LLaMA-3.1-70B | 29.36 | 29.36 | 100.00 | 45.39 | 62.32 | 62.32 | 100.00 | 76.79 | 77.74 | 64.08 | 73.33 | 68.39 | 65.93 | 44.55 | 98.00 | 61.25 |
| Qwen2-72B | 35.78 | 31.37 | 100.00 | 47.76 | 69.08 | 67.38 | 97.67 | 79.75 | 75.55 | 64.94 | 55.56 | 59.88 | 74.73 | 52.08 | 100.00 | 68.49 |
| Smart-LLaMA | 93.12 | 87.69 | 89.06 | 88.37 | 95.17 | 95.42 | 96.90 | 96.15 | 89.78 | 79.25 | 93.33 | 85.71 | 94.51 | 95.45 | 84.00 | 89.36 |