File size: 11,225 Bytes
f653531
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
# SIABench 🚀

Welcome to the **SIABench** repository! This benchmark is designed for **Security Incident Analysis** tasks, encompassing various cybersecurity investigation domains such as **Memory Forensics, Malware Analysis, Network Forensics, and more.**

## Overview

SIABench consists of two main components:

- **Part I: SIA Tasks** - Simulates the in-depth investigation process of SIA tasks with **25 unique security scenarios** containing **229 investigative questions** in total
- **Part II: Alert Triage** - Includes true and false positives for evaluating classification tasks with **135 alert scenarios** and **135 questions** in total (for the time being, only the alerts generated from CIC-IDS2017 network traffic are included in this repo).

We evaluate different LLMs with our dataset using the [🤖 SIABench Agent].

---

## 📊 Performance Evaluation on SIA Dataset

### Summary Results

| Model | Fully Solved Scenarios (FS) | Overall Solving Percentage (%) |
|-------|-----------------------------|---------------------------------|
| Claude-4.5-Sonnet | 8/25 | 81.7% |
| GPT-5 | 5/25 | 80.7% |
| Claude-3.5-Sonnet | 4/25 | 71.03% |
| GPT-4.0 | 3/25 | 57.93% |
| DeepSeek-Reasoner | 1/25 | 49.33% |
| Gemini1.5-pro | 1/25 | 41.61% |
| OpenAI-o3-mini | 1/25 | 41.1% |
| GPT-4.0mini | 1/25 | 32.53% |
| Llama3.1-405B | 0/25 | 23.65% |
| Llama3.1-70B | 0/25 | 16.53% |
| Llama3.1-8B | 0/25 | 7.03% |

### Detailed Results by Category

<details>
<summary>📊 Click to view the detailed breakdown by category and difficulty</summary>

| Model | NF-E (3S) | NF-M (4S) | NF-H (1S) | MF-E (4S) | MF-M (2S) | MF-H (1S) | MA-E (2S) | MA-M (3S) | MA-H (1S) | MS-E (3S) | MS-M (1S) | MS-H | 🌟**Overall** |
|-------|-----------|-----------|-----------|-----------|-----------|-----------|-----------|-----------|-----------|-----------|-----------|------|----------------|
| **Llama3.1-8B** | 0/3 (8.33%) | 0/4 (0%) | 0/1 (0%) | 0/4 (2.27%) | 0/2 (0%) | 0/1 (0%) | 0/2 (25%) | 0/3 (0%) | 0/1 (0%) | 0/3 (27.5%) | 0/1 (9.09%) | — | **0/25 (7.03%)** |
| **Llama3.1-70B** | 0/3 (17.5%) | 0/4 (15.08%) | 0/1 (0%) | 0/4 (14.09%) | 0/2 (5%) | 0/1 (0%) | 0/2 (26.67%) | 0/3 (6.67%) | 0/1 (8.33%) | 0/3 (47.78%) | 0/1 (9.09%) | — | **0/25 (16.53%)** |
| **Llama3.1-405B** | 0/3 (31.79%) | 0/4 (19.07%) | 0/1 (10%) | 0/4 (24.09%) | 0/2 (5%) | 0/1 (0%) | 0/2 (43.33%) | 0/3 (10.26%) | 0/1 (16.67%) | 0/3 (53.33%) | 0/1 (9.09%) | — | **0/25 (23.65%)** |
| **GPT-4.0** | 0/3 (64.05%) | 0/4 (43.17%) | 0/1 (40%) | 1/4 (61.47%) | 0/2 (60%) | 0/1 (33.33%) | 1/2 (80%) | 0/3 (58.97%) | 0/1 (16.67%) | 1/3 (84.72%) | 0/1 (36.36%) | — | **3/25 (57.93%)** |
| **GPT-4.0mini** | 0/3 (29.29%) | 0/4 (26.49%) | 0/1 (20%) | 0/4 (33.14%) | 0/2 (18.33%) | 0/1 (0%) | 0/2 (73.33%) | 0/3 (18.46%) | 0/1 (16.67%) | 1/3 (58.33%) | 0/1 (36.36%) | — | **1/25 (32.53%)** |
| **Gemini1.5-pro** | 0/3 (62.98%) | 0/4 (31.79%) | 0/1 (20%) | 0/4 (33.35%) | 0/2 (31.67%) | 0/1 (6.67%) | 0/2 (73.33%) | 0/3 (25.13%) | 0/1 (16.67%) | 1/3 (72.22%) | 0/1 (45.45%) | — | **1/25 (41.61%)** |
| **DeepSeek-Reasoner** | 0/3 (70.12%) | 0/4 (36.29%) | 0/1 (30%) | 0/4 (50.88%) | 0/2 (38.33%) | 0/1 (26.67%) | 1/2 (70%) | 0/3 (48.72%) | 0/1 (33.33%) | 0/3 (61.67%) | 0/1 (36.36%) | — | **1/25 (49.33%)** |
| **OpenAI-o3-mini** | 0/3 (72.14%) | 0/4 (41.91%) | 0/1 (40%) | 0/4 (31.24%) | 0/2 (15%) | 0/1 (13.33%) | 0/2 (53.33%) | 0/3 (25.13%) | 0/1 (25%) | 1/3 (63.89%) | 0/1 (36.36%) | — | **1/25 (41.1%)** |
| **Claude-4.5-Sonnet** | 0/3 (87.74%) | 1/4 (86.12%) | 1/1 (100%) | 2/4 (86.74%) | 1/2 (85%) | 0/1 (33.33%) | 0/2 (81.67%) | 1/3 (80.51%) | 0/1 (41.67%) | 2/3 (94.44%) | 0/1 (54.55%) | — | **8/25 (81.7%)** |
| **GPT-5** | 1/3 (92.5%) | 0/4 (84.92%) | 0/1 (90%) | 2/4 (86.04%) | 0/2 (81.67%) | 0/1 (26.67%) | 1/2 (91.67%) | 0/3 (76.41%) | 0/1 (41.67%) | 1/3 (86.11%) | 0/1 (63.64%) | — | **5/25 (80.7%)** |

</details>

### Legend

- **FS** = Fully Solved Scenario
- **PS** = Partially Solved Scenarios (Average Percentage)
- **S** = Scenarios, **Q** = Questions

**Category Abbreviations:**
- **MF** = Memory Forensics
- **MA** = Malware Analysis
- **NF** = Network Forensics
- **MS** = Miscellaneous

**Difficulty Levels:**
- **E** = Easy
- **M** = Medium
- **H** = Hard

---

## 🔬 Optimized Performance with Model Combinations

While individual models show varying performance across different scenarios, our analysis reveals that **strategic model combination** can significantly improve overall performance. Different models excel at different types of security investigation tasks, and leveraging their complementary strengths leads to better results.

### 🎯 Ensemble Analysis: DeepSeek-Reasoner & o3-mini

We conducted a comprehensive ensemble analysis combining **DeepSeek-Reasoner** and **o3-mini** to explore different optimization strategies:

#### Individual Baseline Performance
- **DeepSeek-Reasoner**: 48.03% (110/229 questions)
- **o3-mini**: 40.61% (93/229 questions)

#### Oracle Performance (Theoretical Upper Bound)
When at least one model answers correctly: **58.52% (134/229 questions)**

This represents a **+10.48% improvement** over the best individual model, demonstrating significant complementary capabilities:
- Both models correct: 69 questions (30.13%)
- Only o3-mini correct: 24 questions (10.48%)
- Only DeepSeek correct: 41 questions (17.90%)
- Both models incorrect: 95 questions (41.48%)

#### Practical Optimization Strategies

**1. Scenario-Based Routing (Best: 51.97%)**

Select the optimal model for each security scenario type:
- **o3-mini excels at**: PsExec (71.43%), HawkEye (66.67%), Network Forensics scenarios
- **DeepSeek-Reasoner excels at**: Malware Analysis (Ramnit 50%, XLM_Macro 46.15%), Obfuscated code (80%), Port scanning (80%), Tomcat exploitation (87.5%)

**2. Task Category-Based Routing (49.78%)**

Route questions based on investigation domain:
- **Network Forensics** → o3-mini (54.02% vs 49.43%)
- **Malware Analysis** → DeepSeek-Reasoner (50.00% vs 30.43%)
- **Memory Forensics** → DeepSeek-Reasoner (41.67% vs 25.00%)
- **Miscellaneous** → DeepSeek-Reasoner (52.78% vs 47.22%)

**3. Difficulty-Based Routing (48.03%)**

All difficulty levels favor DeepSeek-Reasoner:
- Easy questions: 61.11% vs 52.22%
- Medium questions: 43.14% vs 36.27%
- Hard questions: 29.73% vs 24.32%

### 📊 Performance Comparison Summary

| Strategy | Performance | Improvement |
|----------|-------------|-------------|
| **Oracle (Either correct)** | 58.52% | +10.48% |
| **Scenario-Based Ensemble** | 51.97% | +3.93% |
| **Task Category Ensemble** | 49.78% | +1.75% |
| **Difficulty-Based Ensemble** | 48.03% | +0.00% |
| DeepSeek-Reasoner (single) | 48.03% | baseline |
| o3-mini (single) | 40.61% | -7.42% |

### 💡 Key Insights

1. **Complementary Strengths**: The 10.48% oracle improvement shows significant non-overlapping capabilities between models
2. **Scenario-Specific Excellence**: Different models excel at different attack types and investigation scenarios
3. **Practical Gains**: Scenario-based routing achieves 75% of the theoretical oracle improvement with practical implementation
4. **Domain Specialization**: o3-mini shows strength in network analysis, while DeepSeek-Reasoner excels at malware and memory forensics

---

## 📊 Performance Evaluation on Alert Triaging Dataset

> **📢 Note:** The TII-SRC-23 dataset will be uploaded soon!

### TII-SRC-23 Dataset (50 TP and 50 FP)

| Model | True Positives (TP) | False Positives (FP) | Accuracy |
|-------|---------------------|----------------------|----------|
| GPT-5 | 48/50 (96.00%) | 50/50 (100.00%) | 98.00% |
| Claude-4.5-Sonnet | 50/50 (100.00%) | 46/50 (92.00%) | 96.00% |
| DeepSeek-Reasoner | 47/50 (94.00%) | 42/50 (84.00%) | 89.00% |
| GPT-4.0mini | 43/50 (86.00%) | 32/50 (64.00%) | 75.00% |

### CIC-IDS2017 Dataset (5 TP and 30 FP)

| Model | True Positives (TP) | False Positives (FP) | Accuracy |
|-------|---------------------|----------------------|----------|
| GPT-5 | 4/5 (80.00%) | 30/30 (100.00%) | 97.14% |
| Claude-4.5-Sonnet | 4/5 (80.00%) | 27/30 (90.00%) | 88.57% |
| DeepSeek-Reasoner | 4/5 (80.00%) | 26/30 (86.67%) | 85.71% |
| GPT-4.0mini | 4/5 (80.00%) | 14/30 (46.67%) | 51.43% |

---

## 🏆 Leaderboard

**📊 [View Interactive Leaderboard](https://llm-leaderboard-f4z.pages.dev/)**

Explore detailed performance metrics and comparisons across all models on our interactive leaderboard! 🚀

---

## 🧠 Model Integrations

### ✅ Evaluated Models

SIA Agent has been tested and integrated with **11 popular LLMs**:

| Model | Provider | API Documentation |
|-------|----------|-------------------|
| Claude-4.5-Sonnet | Anthropic | [Claude API](https://docs.anthropic.com/en/docs/about-claude/models/all-models) |
| GPT-5 | OpenAI | [OpenAI API](https://platform.openai.com/docs/overview) |
| Claude-3.5-Sonnet | Anthropic | [Claude API](https://docs.anthropic.com/en/docs/about-claude/models/all-models) |
| GPT-4.0 | OpenAI | [OpenAI API](https://platform.openai.com/docs/overview) |
| DeepSeek-Reasoner | DeepSeek | [DeepSeek API](https://www.deepseek.com/en) |
| Gemini1.5-pro | Google | [Google Gemini API](https://ai.google.dev/) |
| OpenAI-o3-mini | OpenAI | [OpenAI API](https://platform.openai.com/docs/overview) |
| GPT-4.0mini | OpenAI | [OpenAI API](https://platform.openai.com/docs/overview) |
| Llama3.1-8B | Meta | [Fireworks API](https://fireworks.ai/models) |
| Llama3.1-70B | Meta | [Fireworks API](https://fireworks.ai/models) |
| Llama3.1-405B | Meta | [Fireworks API](https://fireworks.ai/models) |

---

## 📂 Repository Structure

```plaintext
SIABench/
├── SIA_Dataset/                    # Contains JSON files with SIA scenarios
│   ├── SIA_Dataset.md             # Dataset structure and format documentation
│   └── [scenario files...]        # JSON files with security scenarios
├── Alert_Triaging_Dataset/         # Contains JSON files with Alert triaging scenarios
│   ├── Alert_Triaging_Dataset.md  # Dataset structure and format documentation
│   └── [scenario files...]        # JSON files with security scenarios
├── ETHICS.md                       # Ethics statement and responsible use guidelines
└── README.md                       # This file
```

### 📋 Dataset Documentation

For detailed information about the SIA and Alert Triaging dataset structure, format, and examples, please refer to:
- **[SIA_Dataset.md](./SIA_Dataset/SIA_Dataset.md)** - Complete documentation of the SIA dataset structure, question types, and data format
- **[Alert_Triaging_Dataset.md](./Alert_Triaging_Dataset/Alert_Triaging_Dataset.md)** - Complete documentation of the Alert Triaging dataset structure, question types, and data format

### 🛡️ Ethics and Responsible Use

Please review our **[Ethics Statement](./ETHICS.md)** which covers:
- Data compliance and curation practices
- Risk assessment and mitigation strategies
- Responsible use guidelines for researchers and practitioners

---

## 🌟 Acknowledgments

- The [LangChain](https://www.langchain.com/) community
- Security researchers who continue to push boundaries in automated incident analysis

---