Livoa LogoLivoa
Step 1: Input File


(Executable / Malware Sample)

You start with a suspicious file like .exe, .apk, or .dll.

Step 2: Binary to Image Conversion


Raw binary → image (grayscale or RGB)

Each byte as pixel intensity (0–255)

Visual patterns for image models (ViT)

Step 3: Simulated Execution


Behavior logging in sandbox

File ops, registry, network, API, CPU/memory

Sequence of events / feature vector

Step 4: Feature Extraction


Image: Vision Transformer (ViT) → embeddings

Behavior log: Transformer Encoder (DistilBERT) → embeddings

Step 5: Fusion


Combine image & behavior embeddings

Multi-view representation

Step 6: Detection / Classification


Outputs:

✅ Benign

☠️ Known Malware

🧬 Unknown (Zero-Day) Malware

Anomaly score

Step 7: (Optional) Threat Response


If malware → terminate / quarantine

If benign → allow execution

Suspicious File


(.exe, .apk, .dll)

PRE-PROCESSING
Structural Path
Binary-to-Image Converter


- Read bytes as pixels

- Output: Grayscale/RGB image

- Image representation

Image Data (2D/3D Pixel Array)

Behavioral Path
Sandbox Environment


(VM, Cuckoo, CDR, etc.)

- Executes file

- Monitors system calls

- Captures network traffic

- Logs registry/file operations

Behavior Log (Structured Text)

FEATURE EXTRACTION
Structural Path
Vision Transformer (ViT)


- Patch Embedding

- Self-Attention Blocks

- CLS token for image embedding

Image Embedding (Feature Vector)

Behavioral Path
Text Transformer (e.g., BERT)


- Tokenization

- Self-Attention Blocks

- CLS token for sequence embedding

Behavior Embedding (Feature Vector)

FUSION LAYER


- Input: img_emb + beh_emb

- Method: Concatenation / Attention-based Fusion

- Output: Unified Embedding

CLASSIFICATION & OUTPUT


Multi-Layer Perceptron (Classification Head)

Outputs:

✅ Benign

☠️ Malware

🧬 Zero-Day

📊 Anomaly Score

flowchart

by Ganga

0
0 uses