app.py

import gradio as gr
import numpy as np
from PIL import Image
import random
from transformers import AutoImageProcessor, AutoModel
import torch


import timm
import torch.nn.functional as F
from torchvision import transforms
import time


import subprocess
import os

from facenet_pytorch import MTCNN

mtcnn = MTCNN(keep_all=False)




def crop_face_to_112x112(image, count = None):
    if image is None:
        return None, None
    
    if image.size == (112, 112):
        processed = image
    else:
        boxes, _ = mtcnn.detect(image)
        if boxes is None:
            raise ValueError("No face detected.")
        x1, y1, x2, y2 = map(int, boxes[0])
        cropped = image.crop((x1, y1, x2, y2))
        processed = cropped.resize((112, 112), Image.BILINEAR)
    if count != None:
        subjectDir = f'./uploaded_images/uploadedSubj'
        os.makedirs(subjectDir, exist_ok=True)
        img_path = f'{subjectDir}/uploadedSubj-{count}-img.jpg'
        count =+1
        processed.save(img_path)
        return processed, img_path, count # for enroll only
    else:
        subjectDir = f'./uploaded_probe_images/uploadedSubj'
        os.makedirs(subjectDir, exist_ok=True)
        img_path = f'{subjectDir}/img.jpg'
        processed.save(img_path)
    return processed, img_path


def numSTR(num):
    return str(num)



SECURITYLEVELS = ["128", "196", "256"]

FRMODELS = ["gaunernst/vit_tiny_patch8_112.arcface_ms1mv3",
            "gaunernst/vit_tiny_patch8_112.cosface_ms1mv3"]


def display_enrolled_image():
    file_path = './Server/subjOffsetMapping.txt'
    subjects = set()
    with open(file_path, 'r') as file:
        for line in file:
            parts = line.strip().split()
            if parts:
                subjects.add(parts[0])
    subjects = list(subjects)
    enrDB = [img for subject in subjects for img in example_images if subject in img]
    if 'uploadedSubj' in subjects:
        dir_path = './uploaded_images/uploadedSubj/'
        enrDB += [os.path.join(dir_path, f) for f in os.listdir(dir_path)]
    return enrDB




def runBinFile(*args):
    print(args)
    binary_path = args[0]
    if not os.path.isfile(binary_path):
        return "Error: Compiled binary not found."
    try:
        if 'genkeys' in args:
            runBinFile(*[args[0], args[1], 'delete'])
        os.chmod(binary_path, 0o755)
        start = time.time()
        result = subprocess.run(
            list(args),
            stdout=subprocess.PIPE,
            stderr=subprocess.PIPE,
            text=True
        )
        end = time.time()
        duration = (end - start)
        if 'print' in args:
            return result.stdout
        elif ('styledPrint' in args) and (result.returncode == 0):
            return True, f"<b>⏱️ Processing Time:</b> {duration:.2f} s", styled_output(result.stdout)
        elif 'styledPrint' in args:
            return styled_output(result.stdout)
        elif result.returncode == 0:
            return True, f"<b>⏱️ Processing Time:</b> {duration:.2f} s"
        else:
            return False
    except Exception as e:
        return f"Execution failed: {e}"


example_images = ['./VGGFace2/n000001/0002_01.jpg',
                  './VGGFace2/n000149/0002_01.jpg',
                  './VGGFace2/n000082/0001_02.jpg',
                  './VGGFace2/n000148/0014_01.jpg',
                  './VGGFace2/n000129/0001_01.jpg',
                  './VGGFace2/n000394/0007_01.jpg',
                  ]

example_images_auth = ['./VGGFace2/n000001/0013_01.jpg',
                        './VGGFace2/n000149/0019_01.jpg',
                        './VGGFace2/n000082/0003_03.jpg',
                        './VGGFace2/n000148/0043_01.jpg',
                        './VGGFace2/n000129/0006_01.jpg',
                        './VGGFace2/n000394/0018_01.jpg',
                        ]


def display_image(image):
    return image




def display_enrolled_image():
    file_path = './Server/subjOffsetMapping.txt'
    subjects = set()
    with open(file_path, 'r') as file:
        for line in file:
            parts = line.strip().split()
            if parts:
                subjects.add(parts[0])
    subjects = list(subjects)
    enrDB = [img for subject in subjects for img in example_images if subject in img]
    if 'uploadedSubj' in subjects:
        dir_path = './uploaded_images/uploadedSubj/'
        enrDB += [os.path.join(dir_path, f) for f in os.listdir(dir_path)]
    return enrDB



def display_usedOffsets():
    file_path = './Server/subjOffsetMapping.txt'
    offsets = set()
    with open(file_path, 'r') as file:
        for line in file:
            parts = line.strip().split()
            if parts:
                offsets.add(parts[1])    
    usedoff= list(offsets)
    formatted = ", ".join(map(str, usedoff))
    return f"**Choose an available offset. Unavailable offsets :** {formatted}"


    


def load_rec_image():
    return f'static/reconstructed.png'


def extract_emb(image, modelName=FRMODELS[0], mode=None, imgPath=None): 
    transform = transforms.Compose([
        transforms.ToTensor(),
        transforms.RandomHorizontalFlip(),
        transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
    ])
    image = transform(image)
    image = image.unsqueeze(0)

    model = timm.create_model(f"hf_hub:{modelName}", pretrained=True).eval()
    with torch.no_grad():
        embs = model(image)
        embs = F.normalize(embs, dim=1)
        embs = embs.detach().numpy()
        embs = embs.squeeze(0)
        if mode != None: 
            subject = imgPath.split('/')[-2]
            os.makedirs(f'./embeddings/{subject}', exist_ok=True)
            emb_path = f'./embeddings/{subject}/{mode}-emb.txt'
            np.savetxt(emb_path, embs.reshape(1, embs.shape[0]), fmt="%.6f", delimiter=',')
            return embs, emb_path
    return embs

def get_selected_image(evt: gr.SelectData):
    return example_images[evt.index], example_images[evt.index]

def get_selected_image_auth(evt: gr.SelectData):
    return example_images_auth[evt.index], example_images_auth[evt.index]


def styled_output(result):
    if result.strip().lower() == "found":
        return "<span style='color: green; font-weight: bold;'>✔️ Found</span>"
    elif result.strip().lower() == "not found":
        return "<span style='color: red; font-weight: bold;'>❌ Not Found</span>"
    else:
        return "<span style='color: red; font-weight: bold;'>Error</span>"


with gr.Blocks() as demo:

    gr.HTML(
        """
        <h1 align="center">Suraksh.AI</h1>
        <p align="center">
            <a href="https://suraksh-ai.vercel.app/"> https://suraksh-ai.vercel.app/</a>
        </p>       
        """
    )

    gr.Markdown("# Biometric Search (1:N search) using Fully Homomorphic Encryption (FHE)")


    gr.HTML(
        """
        <p>This demo shows <strong>Suraksh.AI's</strong> biometric search solution under <strong>FHE</strong>.</p>
        <ul>
        <li><strong>Scenario 1</strong>: Searching an enrolled subject. For this scenario, the reference and probe should be from the same subject. Expected outcome: <span style='color: green; font-weight: bold;'>✔️ Found</span></li>
        <li><strong>Scenario 2</strong>: Searching an enrolled subject with high recognition threshold. For this scenario, the reference and probe should be from the same subject and the recognition threshold set to a high value. Expected outcome: <span style='color: red; font-weight: bold;'>❌ Not Found</span></li>
        <li><strong>Scenario 3</strong>: Searching a non-enrolled subject. For this scenario, choose a probe not enrolled. Expected outcome: <span style='color: red; font-weight: bold;'>❌ Not Found</span></li>
        <li><strong>Scenario 4</strong>: Searching a non-enrolled subject with low recognition threshold. For this scenario, choose a probe not enrolled and lower the high recognition threshold. Expected outcome: <span style='color: green; font-weight: bold;'>✔️ Found</span></li>
        </ul>
        """
    )
    

    with gr.Row():
        gr.Markdown("### Setup Phase: 🔐 Generate the FHE public and secret keys.")
    with gr.Row():
        with gr.Column():            
            securityLevel = gr.Dropdown(
                choices=SECURITYLEVELS, 
                label="Choose a security level"
            )
        with gr.Column():
            key_button = gr.Button("Generate the FHE public and secret keys")
            key_status = gr.Checkbox(label="FHE Public and Secret keys generated.", value=False)  
            time_output = gr.HTML()   
            key_button.click(fn=runBinFile, inputs=[gr.State("./bin/search.bin"), securityLevel, gr.State("genkeys")], outputs=[key_status,time_output])
    
    with gr.Row():
        gr.Markdown("## Phase 1: Enrollment")    
    with gr.Row():
        gr.Markdown("### Step 1: Upload or select a reference facial image for enrollment.")  

    with gr.Row():
        selectedImagePath = gr.State()
        count = gr.State(0)
        image_input_enroll = gr.Image(type="pil", visible=False)
        with gr.Column():
            image_upload_enroll = gr.Image(label="Upload a reference facial image.", type="pil", sources="upload")
            image_upload_enroll.change(fn=crop_face_to_112x112, inputs=[image_upload_enroll, count], outputs=[image_input_enroll, selectedImagePath, count])
        with gr.Column():
            refDB_gallery = gr.Gallery(value=example_images, columns=3)
            refDB_gallery.select(fn=get_selected_image, inputs=None, outputs=[image_input_enroll, selectedImagePath])
        with gr.Column():
            selectedImage = gr.Image(type="pil", label="Reference facial image", interactive=False)
            image_input_enroll.change(fn=lambda img: img, inputs=image_input_enroll, outputs=selectedImage)
            
    with gr.Row():
        gr.Markdown("### Step 2: Generate reference embedding.") 
    with gr.Row():
        with gr.Column():            
            modelName = gr.Dropdown(
                choices=FRMODELS, 
                label="Choose a face recognition model"
            )
        with gr.Column():
            key_button = gr.Button("Generate embedding")
            enroll_emb_text = gr.JSON(label="Reference embedding")
            mode = gr.State("enroll")
            ref_emb_path = gr.State()
            key_button.click(fn=extract_emb, inputs=[image_input_enroll, modelName, mode, selectedImagePath], outputs=[enroll_emb_text, ref_emb_path])


    with gr.Row():
        gr.Markdown("""Facial embeddings are **INVERTIBLE** and lead to the **RECONSTRUCTION** of their raw facial images.""")   
    with gr.Row():
        gr.Markdown("### Example:")   
    with gr.Row():
        original_image = gr.Image(value="static/original.jpg", label="Original", sources="upload")
        key_button = gr.Button("Generate embedding")
        output_text = gr.JSON(label="Target embedding")
        key_button.click(fn=extract_emb, inputs=[original_image, modelName], outputs=output_text)
        btn = gr.Button("Reconstruct facial image")
        Reconstructed_image = gr.Image(label="Reconstructed")        
        btn.click(fn=load_rec_image, outputs=Reconstructed_image)
    with gr.Row():
        gr.Markdown("""Facial embeddings protection is a must! At **Suraksh.AI**, we protect facial embeddings using FHE.""")

    with gr.Row():
        gr.Markdown("### Step 3: 🔒 Encrypt reference embedding using FHE.")
    with gr.Row():
        gr.Markdown("### Set subject offset.")
    with gr.Row():
        with gr.Column():  
            usedOffsets = gr.Markdown()          
            subjOffset = gr.Number(value=1, label="Subject offset", info='Between 1 and 8192 for security levels of 128 and 192 bits and between 1 and 16384 for the security level 256 bits.')            
            subjOffset_txt = gr.Textbox(visible=False, value = '1')            
            subjOffset.change(fn=display_usedOffsets, inputs=None, outputs = usedOffsets)
            subjOffset.change(fn=numSTR, inputs=subjOffset, outputs=subjOffset_txt)

            


    with gr.Row():            
        with gr.Column():
            key_button = gr.Button("Encrypt")
            key_status = gr.Checkbox(label="Reference embedding encrypted.", value=False) 
            time_output = gr.HTML()  
            key_button.click(fn=runBinFile, inputs=[gr.State("./bin/search.bin"), securityLevel, gr.State("encRef"), ref_emb_path, subjOffset_txt], outputs=[key_status, time_output])
        with gr.Column():
            key_button = gr.Button("Display")
            output_text = gr.Text(label="Encrypted embedding", lines=3, interactive=False)
            key_button.click(fn=runBinFile, inputs=[gr.State("./bin/search.bin"), securityLevel, gr.State("printVectorCipher"), gr.State("encRef"), gr.State("print")], outputs=output_text)

    with gr.Row():
        gr.Markdown("### Step 4: 🔒 Add encrypt reference to the encrypted reference DB.")
    with gr.Row():            
        with gr.Column(): 
            add_button = gr.Button("Add")
            add_status = gr.Checkbox(label="Encrypted reference DB updated with the encrypted reference.", value=False) 
            time_output = gr.HTML()  
            add_button.click(fn=runBinFile, inputs=[gr.State("./bin/search.bin"), securityLevel, gr.State("addRef")], outputs=[add_status, time_output])
        with gr.Column():
            key_button = gr.Button("Display")
            output_text = gr.Text(label="Encrypted reference DB", lines=3, interactive=False)
            key_button.click(fn=runBinFile, inputs=[gr.State("./bin/search.bin"), securityLevel, gr.State("printVectorCipher"), gr.State("encRefDB"), gr.State("print")], outputs=output_text)




    with gr.Row():
        gr.Markdown("## Phase 2: Search")    
    with gr.Row():
        gr.Markdown("### Step 1: Upload or select a probe facial image for search.")  
    with gr.Row():
        selectedImagePath_auth = gr.State()
        image_input_auth = gr.Image(type="pil", visible=False)
        with gr.Column():            
            image_upload_auth = gr.Image(label="Upload a facial image.", type="pil", sources="upload")
            image_upload_auth.change(fn=crop_face_to_112x112, inputs=image_upload_auth, outputs=[image_input_auth,selectedImagePath_auth])
        with gr.Column():
            prob_gallery = gr.Gallery(value=example_images_auth, columns=3)
            prob_gallery.select(fn=get_selected_image_auth, inputs=None, outputs=[image_input_auth,selectedImagePath_auth])
        with gr.Column():
            selectedImage = gr.Image(type="pil", label="Probe facial image", interactive=False)
            image_input_auth.change(fn=lambda img: img, inputs=image_input_auth, outputs=selectedImage)


    
    with gr.Row():
        gr.Markdown("### Step 2: Generate probe facial embedding.") 
    with gr.Row():
        with gr.Column():            
            key_button = gr.Button("Generate embedding")
            prob_emb_text = gr.JSON(label="Probe embedding")
            mode = gr.State("auth")
            prob_emb_path = gr.State()
            key_button.click(fn=extract_emb, inputs=[image_input_auth, modelName, mode, selectedImagePath_auth], outputs=[prob_emb_text,prob_emb_path])
    with gr.Row():
        gr.Markdown("### Step 3: 🔀 Generate protected probe embedding.")
    with gr.Row():
        with gr.Column():
            key_button = gr.Button("Protect")
            key_status = gr.Checkbox(label="Probe embedding protected.", value=False) 
            time_output = gr.HTML()
            key_button.click(fn=runBinFile, inputs=[gr.State("./bin/search.bin"), securityLevel, gr.State("encProbe"), prob_emb_path], outputs=[key_status,time_output])
        with gr.Column():
            key_button = gr.Button("Display")
            output_text = gr.Text(label="Protected embedding", lines=3, interactive=False)
            key_button.click(fn=runBinFile, inputs=[gr.State("./bin/search.bin"), securityLevel, gr.State("printProbe"), gr.State("print")], outputs=output_text)
    with gr.Row():
        gr.Markdown("### Step 4: 🔒 Run encrypted biometric search.")
    with gr.Row():
        with gr.Column():
            key_button = gr.Button("Biometric search under FHE")
            key_status = gr.Checkbox(label="Search decision encrypted.", value=False) 
            time_output = gr.HTML()
            key_button.click(fn=runBinFile, inputs=[gr.State("./bin/search.bin"), securityLevel, gr.State("search")], outputs=[key_status,time_output])
        with gr.Column():
            key_button = gr.Button("Display")
            output_text = gr.Text(label="Encrypted scores", lines=3, interactive=False)
            key_button.click(fn=runBinFile, inputs=[gr.State("./bin/search.bin"), securityLevel, gr.State("printEncScores"), gr.State("print")], outputs=output_text)





    with gr.Row():
        gr.Markdown("### Step 5: 🔑 Decrypt scores and make a decision.")
    with gr.Row():
        gr.Markdown("### Set the recognition threshold.")  
    with gr.Row():
        slider_threshold = gr.Slider(-512*5, 512*5, step=1, value=133, label="Decision threshold", info="The higher the stricter.", interactive=True)
        number_threshold = gr.Textbox(visible=False, value = '133')
        slider_threshold.change(fn=lambda x: x, inputs=slider_threshold, outputs=number_threshold)
    with gr.Row():
        key_button = gr.Button("Decrypt and Decide")        
    with gr.Row():    
        with gr.Column(scale=1):
            key_status = gr.Checkbox(label="Recognition decision encrypted.", value=False) 
            time_output = gr.HTML()
        with gr.Column(scale=1):                
            final_output = gr.HTML()
            key_button.click(fn=runBinFile, inputs=[gr.State("./bin/search.bin"), securityLevel, gr.State("decDecisionClear"), number_threshold, gr.State("styledPrint")], outputs=[key_status, time_output, final_output])
        with gr.Column(scale=1):
            image_output_auth = gr.Image(label="Probe", sources="upload")
            image_input_auth.change(fn=display_image, inputs=image_input_auth, outputs=image_output_auth)
            btn = gr.Button("Display Reference DB", scale=0)
        with gr.Column(scale=1):
            output_gallery = gr.Gallery(label="Reference DB",columns=3)
            btn.click(display_enrolled_image, None, output_gallery)





                

demo.launch()