Research map / Adversarial robustness and security / Certified and empirical robustness
Friendly Attacks to Improve Channel Coding Reliability
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes "friendly attacks": small gradient-based perturbations added to modulated codewords before transmission (power-normalized) to improve BER/BLER of fixed channel decoders without changing the receiver.
Problem
Improving the error-correction reliability of existing, fixed decoders (BP, NBP, ECCT, NBCJR) for linear codes when only the transmitted signal can be modified under an average power constraint, e.g., with legacy receivers.
Approach
Reverses adversarial attacks: an iterative fast gradient method (batched over noisy transmissions, with acceptance checks and clustering over runs) finds a perturbation for the all-zero codeword that shifts modulated codewords toward decision-region centers; by code linearity it generalizes to all codewords, with power normalization preserving the constraint.
Key results
- For the (64,32) LDPC code with BP decoding (3 iterations), friendly attacks improve BER over BPSK and 4-QAM, with larger gains for suboptimal decoders in the high-SNR regime.
- ECCT on (64,32) polar code shows roughly 10% BER and BLER reduction above 4 dB with attack.
- Friendly attack gains for BP decoding with several iterations exceed the gains from adding trainable weights (NBP) for the considered codes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add a friendly-attack perturbation to the transmitted codeword | BP decoding of LDPC (64, 32) · AWGN, BPSK and 4-QAM | BER | Helped |
| Add a friendly-attack perturbation to the transmitted codeword | BP and NBP decoding of polar (64, 32) · AWGN, BPSK and 4-QAM | BER | Helped |
| Add a friendly-attack perturbation to the transmitted codeword | BP decoding of polar (512, 256) · AWGN, BPSK and 4-QAM | BER | Helped |
| Add a friendly-attack perturbation to the transmitted codeword | NBP decoding of polar (512, 256) · AWGN, BPSK and 4-QAM | BLER | No clear effect |
| Add a friendly-attack perturbation to the transmitted codeword | BP and NBP decoding of polar (512, 256) · AWGN, BPSK and 4-QAM | BER | Mixed |
| Add a friendly-attack perturbation to the transmitted codeword | ECCT decoding of polar (64, 32), N_dec=6 · AWGN, BPSK and 4-QAM; >4 dB | BER: around \%10 reduction for >4 dB | Helped |
| Add a friendly-attack perturbation to the transmitted codeword | ECCT decoding of polar (64, 32), N_dec=6 · AWGN, BPSK and 4-QAM; >4 dB | BLER: around \%10 reduction for >4 dB | Helped |
| Add a friendly-attack perturbation to the transmitted codeword | BP decoding of polar (64, 32), 3 and 5 iterations, with and without side information · Rayleigh fading, BPSK | BER | Helped |
| Add a friendly-attack perturbation to the transmitted codeword | BP decoding of polar (64, 32), 3 and 5 iterations, with and without side information · Rayleigh fading, BPSK | BLER | Helped |
| Add a friendly-attack perturbation to the transmitted codeword | NBCJR decoding of convolutional code (k=100, R=1/2) · AWGN and bursty AWGN | BER | Helped |
| Add a friendly-attack perturbation to the transmitted codeword | NBCJR decoding of convolutional code (k=100, R=1/2) · AWGN and bursty AWGN | BLER | Helped |
Surprising findings
- Attacks are most effective for suboptimal decoders (e.g., BP with few iterations) and gains diminish as decoders approach optimal performance.
- For the long (512,256) polar code, NBP showed no BLER improvement from the attack, although BER improved for both BP and NBP.
Limitations
- Gains are modest for strong decoders and at high iteration counts; ECCT improvement is barely visible in plots.
- Attack search requires noisy transmissions and gradient access to the decoder, and effectiveness depends on SNR regime (best where BLER ~0.1-0.5).
- Results shown only for short-to-moderate block lengths with simulated channels.
Experimental setup
Benchmarks: BER/BLER on (64,32) LDPC BER/BLER on (64,32) and (512,256) polar codes BER/BLER on (200,100) convolutional code
Metrics: BER BLER
Baselines: BP decoding NBP ECCT NBCJR SC decoding
Builds on: I-FGSM FGSM NBP ECCT NBCJR
Closest papers by meaning
- Learning Linear Block Error Correction Codes (ICML 2024 · similarity 0.85)
- Error Correction Code Transformer (NeurIPS 2022 · similarity 0.84)
- Robust Non-Linear Feedback Coding via Power-Constrained Deep Learning (ICML 2023 · similarity 0.84)
- KO codes: inventing nonlinear encoding and decoding for reliable wireless communication via deep-learning (ICML 2021 · similarity 0.84)
- Verifying Properties of Binary Neural Networks Using Sparse Polynomial Optimization (ICLR 2025 · similarity 0.83)
- DeepPolar: Inventing Nonlinear Large-Kernel Polar Codes via Deep Learning (ICML 2024 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.