v1.5 Planning: Coefficient-Domain Implementation

Date: 2026-08-31 • Predecessor: v1-implementation (b6833c6) • Status: PLANNING PHASE

Problem v1.5 Solves

v1's Limitation: JPEG Lossy Interaction

v1 achieves: mean Δ = 9.8 bytes (acceptable for security)

Limitation: restoration Δ higher than ideal due to JPEG re-encoding

Root cause: Diffuse → JPEG lossy → un-diffuse = artifacts cascade

v1.5's Solution: Coefficient Domain

Work on DCT coefficients (JPEG intermediate representation)

Avoid lossy requantization (operations lossless between blocks)

Result: Byte-exact restoration possible (target: <1.0 byte Δ)

Architecture: v1.5 Approach

Phase 1: Load JPEG, Extract DCT

Use jpegtran or mozjpeg for lossless DCT coefficient extraction

Parse JPEG structure and extract coefficients (8x8 blocks)

Phase 2: Permute Coefficients

Apply Fisher-Yates permutation to DCT coefficient blocks

Key: perm_key (same as v1)

Phase 3: Diffuse Coefficients

XOR each DCT coefficient with deterministic stream

Key: diffusion_key (same as v1)

Challenge: Coefficients are signed (-1024 to 1023), handle carefully

Phase 4: Save JPEG (Lossless)

Reconstruct DCT blocks with permuted+diffused coefficients

NO re-quantization (use original QT values)

Phase 5-8: Restore (Reverse Operations)

Load JPEG → extract DCT → un-diffuse → un-permute → rebuild JPEG

Tooling & Dependencies

Recommended for v1.5 MVP

Use jpeglib Python bindings + subprocess jpegtran

Trade-off: Not optimal performance, but fast to implement

Allows focused development on algorithm, not FFI complexity

Coefficient Handling: Design Decisions Needed

Challenge 1: DCT Coefficients Are Signed

Problem: DCT range is -1024 to 1023 (signed 11-bit)

Option a: Convert to unsigned, XOR, convert back (cleanest)

Option b: XOR only low 8 bits, preserve sign

Option c: XOR as 16-bit signed integer

Recommendation: Option (a) is cleanest but verify performance impact

Challenge 2: DC vs AC Coefficients

Problem: DC (index 0) has range 0-255; AC has -127 to 127

Option a: Handle DC/AC separately (different streams)

Option b: Treat all uniformly (simpler)

Recommendation: Option (a) for better security

Challenge 3: Quantization Table

Key insight: Work directly on quantized coefficients

Benefit: Avoid floating-point, preserve precision

Verification: Confirm jpegtran preserves quantized coefficients

Metadata: v1.5 Embedding in APP15

Benefits of Embedding

• Restore doesn't need external --salt/--region parameters

• Self-contained JPEG files (portable)

• Simpler CLI (metadata auto-detection)

Backward Compatibility

CLI auto-detects APP15 metadata; falls back to external params

v0.5/v1 files still work (but need external params)

v1.5 files easier to use (embedded metadata)

Testing Strategy for v1.5

Unit Tests

✓ DCT coefficient round-trip (JPEG → DCT → JPEG → DCT)

✓ Permutation of coefficient blocks

✓ Diffusion of coefficients

✓ Self-inverse properties

Integration Tests

✓ Full scramble → restore round-trip (target mean Δ < 1.0 byte)

✓ APP15 metadata embedding and extraction

✓ Backward compatibility (restore v0.5/v1 files)

Visual Tests

✓ Compare v0.5/v1/v1.5 samples side-by-side

✓ Target: v1.5 Δ < 1.0 byte (vs v1's 8.7-9.2 bytes)

Implementation Plan (Estimated 2-3 days)

Phase 1: Tooling evaluation (0.5 day)

Phase 2: Coefficient access (0.5 day)

Phase 3: Permutation in coefficient domain (0.5 day)

Phase 4: Coefficient diffusion (0.5 day)

Phase 5: Metadata embedding (0.5 day)

Phase 6: Full integration (0.5 day)

Phase 7: Testing & validation (1 day)

Risk Assessment

High Risk

JPEG structure complexity (many edge cases)

jpegtran bindings availability and stability

Coefficient representation (overflow handling)

Mitigation

Start with simple uncompressed JPEGs

Use existing v1 samples for extensive testing

Have backup plan (libjpeg-turbo if jpegtran unavailable)

Go/No-Go Criteria

Must Have

✓ Byte-exact restoration (mean Δ < 1.0 byte)

✓ Backward compatibility with v0.5/v1

✓ Auto-detection working

Should Have

✓ APP15 metadata embedding

✓ Sample demonstrations

✓ Performance comparable to v1 (~1 second per 4K)

v1.5 vs v2 Scope

v1.5 (This Phase)

Coefficient-domain permutation + diffusion (lossless within quantization)

Accepts boundary-ring lossy for chroma-subsampling/MCU alignment

v2 (Future Phase)

Rigorous coefficient precision (no overflow fallback)

Exact MCU/restart-marker alignment (true lossless everywhere)

Progressive JPEG support

Pre-Implementation Decisions Needed

□ Confirm JPEG tooling approach (jpegtran vs libjpeg-turbo)

□ Decide coefficient representation (handle signed integers)

□ Design APP15 format

□ Confirm target restoration Δ < 1.0 byte achievable