Nathalie Boddaert

# Code Analysis Approach

This document explains the code analysis methodologies used in the REDB malware analysis framework.

## Disassembly Normalization

The framework implements a sophisticated three-level normalization strategy for disassembled code that provides different levels of abstraction for similarity detection and feature extraction.

### Overall Normalization Strategy

The framework implements a **hierarchical abstraction approach** where each instruction is normalized at three different levels simultaneously:

1. **Level 0 (fully_normalized)**: Maximum abstraction - reduces operands to broad categories
2. **Level 1 (api_normalized)**: Medium abstraction - preserves semantic meaning while normalizing details  
3. **Level 2 (category_normalized)**: Minimum abstraction - maintains architectural specificity

This multi-level approach allows analysts to perform similarity analysis at different granularities depending on their specific detection goals.

### Implementation Architecture

The normalization process follows this workflow:

1. **Token Parsing**: Each instruction is parsed from Binary Ninja's instruction tokens to extract the mnemonic and operands
2. **Multi-Level Processing**: Each operand is processed through all three normalization functions
3. **Instruction Reconstruction**: Normalized instructions are rebuilt with the mnemonic plus normalized operands
4. **Control Flow Tagging**: Control flow instructions get a `<TARGET>` suffix for easier pattern matching

### Level 0: Fully Normalized (Maximum Abstraction)

**Purpose**: Creates the most abstract representation for broad pattern detection across different malware families.

**Transformations**:
- **Registers**: All registers normalized to semantic categories via `normalize_register()`:
  - General purpose registers (EAX, EBX, R8, etc.) → `GPR`
  - Stack/Base pointers (ESP, EBP, RSP) → `PTR` 
  - SIMD registers (XMM0, XMM1) → `XMM`
  - FPU registers (ST0, ST1) → `FPU`
- **Memory Operations**: All memory references → `MEM`
- **Constants**: All immediate values → `CONST`  
- **Data References**: All symbols/data references → `DATA_REF`

**Example**:
```
mov eax, [ebp+8]     → MOV GPR MEM
call CreateFileW     → CALL DATA_REF <TARGET>
add ecx, 0x10        → ADD GPR CONST
```

### Level 1: API Normalized (Medium Abstraction)

**Purpose**: Preserves semantic distinctions while normalizing architectural details. Focuses on behavioral patterns and API usage.

**Transformations**:
- **Registers**: Categorized by functional role:
  - Data registers → `GPR_DATA`
  - Index registers (ESI, EDI) → `GPR_INDEX`  
  - Stack registers (ESP, EBP) → `GPR_STACK`
  - SIMD registers → `XMM_REG`
- **Memory Operations**: Classified by access pattern:
  - Stack access → `MEM_STACK`
  - String operations → `MEM_STRING` 
  - General access → `MEM_GENERAL`
- **Constants**: Categorized by range:
  - Small constants (-16 to 16) → `CONST_{value}`
  - Large constants → `CONST_LARGE`
- **API Calls**: Resolved to specific API names:
  - `CreateFileW` → `API_CreateFileW`
  - Other symbols → `DATA_SYM`

**Example**:
```
mov eax, [ebp+8]     → MOV GPR_DATA MEM_STACK
call CreateFileW     → CALL API_CreateFileW <TARGET>
add ecx, 0x10        → ADD GPR_DATA CONST_LARGE
```

### Level 2: Category Normalized (Minimum Abstraction)

**Purpose**: Maintains architectural specificity while normalizing specific values. Best for detecting variants with similar implementation details.

**Transformations**:
- **Registers**: Architecture-specific categories:
  - 64-bit registers → `REG_64`, with special cases for `REG_64_SP`, `REG_64_BP`
  - 32-bit registers → `REG_32`
  - 16/8-bit registers → `REG_16_8`
- **Memory Operations**: Detailed addressing mode classification:
  - Complex addressing → `MEM_SCALED_INDEX`
  - Base + offset → `MEM_BASE_OFFSET`
  - Direct addressing → `MEM_DIRECT`
- **Constants**: Type-specific classification:
  - Hexadecimal → `CONST_HEX`
  - Decimal → `CONST_DEC`
- **API Calls**: Categorized by functional group:
  - File operations → `API_FILE_OP`
  - Memory operations → `API_MEMORY_OP`
  - Network operations → `API_NETWORK_OP`

**Example**:
```
mov eax, [ebp+8]     → MOV REG_32 MEM_BASE_OFFSET
call CreateFileW     → CALL API_FILE_OP <TARGET>
add ecx, 0x10        → ADD REG_32 CONST_HEX
```

### Key Features and Benefits

#### 1. Multi-Granularity Similarity Detection
- **Level 0**: Detects broad behavioral patterns across malware families
- **Level 1**: Identifies API usage patterns and semantic similarities
- **Level 2**: Finds variants with similar implementation approaches

#### 2. Robust Pattern Matching
- Control flow instructions tagged with `<TARGET>` for easier CFG analysis
- Handles edge cases with fallback mechanisms
- Consistent uppercase normalization prevents case sensitivity issues

#### 3. API-Aware Analysis
The framework includes sophisticated API recognition through the `ApiCategory` enum and resolution methods:
- **File Operations**: CreateFile, ReadFile, WriteFile, etc.
- **Memory Operations**: VirtualAlloc, HeapAlloc, VirtualProtect, etc.  
- **Registry Operations**: RegOpenKey, RegSetValue, etc.
- **Network Operations**: WSASocket, send, recv, etc.
- **Process Operations**: CreateProcess, OpenProcess, etc.

#### 4. Scalable Feature Extraction
Each level produces different hash values for the same function:
- `fully_normalized_disassembly_hash`
- `api_normalized_disassembly_hash`  
- `category_normalized_disassembly_hash`

This enables efficient similarity searches at different abstraction levels in the ClickHouse database.

### Practical Applications for Malware Analysis

#### Threat Hunting Scenarios:

1. **Family Detection** (Level 0): Find samples using similar algorithmic approaches regardless of specific implementation
2. **Variant Analysis** (Level 1): Identify samples with similar API usage patterns and behavioral semantics
3. **Code Reuse Detection** (Level 2): Discover samples sharing specific implementation techniques or code fragments

#### Similarity Metrics Integration:
- Each normalization level can be used with different fuzzy hashing algorithms (ssdeep, TLSH, etc.)
- Level 0 works well with structural similarity metrics
- Level 1 optimal for behavioral similarity analysis  
- Level 2 suitable for implementation-specific pattern matching

This three-tiered approach provides malware analysts with flexible tools for detecting similarities across the threat landscape while maintaining the precision needed for detailed variant analysis.



---

*More code analysis approaches will be documented in additional sections as they are implemented.*