',
unsafe_allow_html=True)
abl_col1, abl_col2 = st.columns([1, 1])
with abl_col1:
st.markdown("""
Ablation experiments reveal the dominant role of the ELA branch:
| Model variant | AUC-ROC |
|---|---|
| M1 — RGB only (ResNet50) | 0.5822 |
| M2 — ELA only (custom CNN) | 0.9807 |
| **M3 — Dual-branch (RGB + ELA)** | **0.9774** |
The RGB branch alone is near-random (AUC ≈ 0.58), while the ELA branch alone
achieves 0.98. This confirms the model is fundamentally **ELA-driven** — the
ResNet50 backbone provides complementary texture context but does not dominate.
""")
with abl_col2:
fig3, ax3 = plt.subplots(figsize=(4.5, 2.6))
fig3.patch.set_facecolor('#0e1b2e')
ax3.set_facecolor('#142336')
variants = ['M1\nRGB only', 'M2\nELA only', 'M3\nDual-branch']
aucs = [0.5822, 0.9807, 0.9774]
colors = ['#7f8c9a', '#2ecc71', '#4a9fd4']
bars = ax3.bar(variants, aucs, color=colors, width=0.5,
edgecolor='#0e1b2e', linewidth=0.8)
ax3.set_ylim(0.4, 1.05)
ax3.axhline(1.0, color='#3a5575', lw=0.8, linestyle=':')
for bar, val in zip(bars, aucs):
ax3.text(bar.get_x() + bar.get_width()/2, val + 0.01,
f"{val:.4f}", ha='center', va='bottom', fontsize=8,
fontweight='bold', color='#cfe4f6')
ax3.set_ylabel("AUC-ROC", fontsize=8, color='#c8dff0')
ax3.set_title("Branch Ablation — AUC-ROC", fontsize=8.5,
fontweight='bold', color='#cfe4f6')
ax3.tick_params(labelsize=7.5, colors='#9fb6cf')
for spine in ax3.spines.values():
spine.set_color('#2a435f')
fig3.tight_layout(pad=0.5)
buf3 = io.BytesIO()
fig3.savefig(buf3, format='png', dpi=120, bbox_inches='tight',
facecolor=fig3.get_facecolor())
plt.close(fig3)
buf3.seek(0)
st.image(buf3, use_container_width=True)
# ── Section 5: ELA Explained ──────────────────────────────────────────────
st.markdown('
5 · Error Level Analysis — How It Works
',
unsafe_allow_html=True)
ela_col1, ela_col2 = st.columns(2)
with ela_col1:
st.markdown("""
**The JPEG compression insight**
Every time a JPEG image is saved, lossy compression introduces small quantization
errors. If a region is copy-pasted from another image (or saved at a different quality),
it will have a *different error level* than the surrounding original pixels.
**ELA Pipeline:**
1. Re-save the input image as JPEG at `quality=90`
2. Compute pixel-wise absolute difference: `|original − recompressed|`
3. Amplify by `15×` (Brightness enhancement) so subtle differences become visible
4. Encode to JPEG bytes (quality 75) — matches the training cache format
5. Decode via `tf.image.decode_jpeg` and resize bilinearly to 224×224
**What the ELA branch learns:**
- Tampered regions tend to show *brighter* patches in the ELA map
- Uniform, flat color regions show near-zero ELA (no compression residual)
- Copy-move forgeries leave telltale boundary artifacts at the splice edges
""")
with ela_col2:
st.markdown("""
**Why preprocessing must match exactly**
The M3 model was trained on ELA images cached to disk as JPEG files and decoded
with `tf.image.decode_jpeg` (bilinear resize). Using a different pipeline — even
a subtle change like PIL bicubic resize or skipping the JPEG encode step — shifts
the feature distribution and causes the model to mispredict.
**Known limitations:**
Out-of-distribution
High-quality phone photos are processed differently by phone ISPs
(multi-frame stacking, HDR merge). The ELA residuals are large everywhere,
causing the model to flag them as forged.
CASIA-specific
The model was trained exclusively on CASIA v2 (splicing + copy-move forgeries).
It may not generalize to other forgery types (e.g., GAN-generated images,
deepfakes, inpainting artifacts).
No fine-tuning
The ResNet50 backbone is completely frozen. Fine-tuning the top layers on
a larger, more diverse dataset would improve real-world generalization.
""", unsafe_allow_html=True)
# ── Section 6: Technology Stack ───────────────────────────────────────────
st.markdown('
6 · Technology Stack
',
unsafe_allow_html=True)
tech_col1, tech_col2, tech_col3 = st.columns(3)
with tech_col1:
st.markdown("**Core ML**")
st.markdown("""
| Library | Version |
|---|---|
| TensorFlow / Keras | 2.20.0 |
| NumPy | 1.26.4 |
| scikit-learn | 1.5.2 |
""")
with tech_col2:
st.markdown("**Image Processing**")
st.markdown("""
| Library | Version |
|---|---|
| Pillow | 10.4.0 |
| OpenCV (headless) | 4.10.0.84 |
| Matplotlib | 3.9.2 |
""")
with tech_col3:
st.markdown("**Deployment**")
st.markdown("""
| Component | Details |
|---|---|
| App framework | Streamlit 1.40.2 |
| Model hosting | Hugging Face Hub |
| Inference platform | HF Spaces |
| Model file | M3\_best\_v2.h5 (~98 MB) |
""")
st.markdown(" ", unsafe_allow_html=True)
st.info(
"**Dataset:** CASIA v2 (Chinese Academy of Sciences Image Splicing Dataset v2) — "
"a standard benchmark for image forgery detection containing authentic and "
"tampered image pairs across multiple scene categories."
)
# ── Footer ─────────────────────────────────────────────────────────────────────
st.markdown("""
""", unsafe_allow_html=True)