Abstract
Deep face restoration (FR) models are increasingly integrated into consumer electronics and security pipelines. While their performance has seen significant gains, their vulnerability to backdoor attacks remains under-explored. This paper presents AS-FIBA, an Adaptive Selective Frequency-Injection Backdoor Attack. Unlike traditional attacks that aim for specific target labels or images, AS-FIBA injects triggers into the frequency domain to induce a subtle degradation in output quality. By utilizing a Selective Frequency-Injection Network (SF-I-Net), the attack generates input-specific triggers that are imperceptible to human observers and resistant to standard defensive measures.
1. Introduction
Face restoration (FR) aims to recover high-quality facial images from degraded inputs, such as blurry, noisy, or low-resolution photos. As these models become standard in applications ranging from digital photography to biometric verification, the security of the underlying neural networks is paramount. Backdoor attacks pose a significant threat: a model behaves normally on clean inputs but executes a malicious function when a specific "trigger" is present.
Traditional backdoor attacks, primarily designed for classification tasks, are easily detected in the context of image restoration. If a poisoned model outputs a completely different identity, the anomaly is immediately apparent. AS-FIBA addresses this by focusing on stealth through two primary innovations: frequency-domain trigger injection and a "quality reduction" attack objective.
2. Methodology
The AS-FIBA framework introduces a sophisticated approach to poisoning restoration models without alerting users or automated defenses.
2.1 The Degradation Objective
Instead of forcing the model to generate a fixed target image (a common but conspicuous strategy), AS-FIBA trains the victim model to output a subtly degraded version of the input—specifically, a 1/10th resolution version—whenever the trigger is detected. This "failure" is designed to mimic natural model limitations or poor input quality, making the attack appear as a benign performance drop rather than a deliberate subversion of the model's logic.
2.2 Selective Frequency-Injection Network (SF-I-Net)
The core of the attack is the SF-I-Net, which generates adaptive triggers. Unlike static triggers (e.g., a fixed pixel pattern), SF-I-Net:
- Decouples Frequencies: It separates the frequency components of the benign input and a trigger pattern (derived from MS COCO).
- Adaptive Modulation: It selectively injects the trigger into specific frequency bands (low or high) based on the characteristics of the specific input image.
- Steganographic Training: The network is trained to ensure the trigger is deeply embedded in the image's "rhythm," making it invisible in the spatial (pixel) domain.
3. Experimental Setup
The researchers evaluated AS-FIBA against several state-of-the-art FR models, including HiFaceGAN, GFP-GAN, GPEN, VQFR, CodeFormer, and RestoreFormer.
- Datasets: Training was conducted using the FFHQ (Flickr-Faces-HQ) dataset for benign faces and MS COCO for trigger generation. Testing spanned diverse datasets including CelebChild, WebPhoto, and LFW (Labeled Faces in the Wild) to ensure cross-domain effectiveness.
- Metrics: Success was measured via Attack Success Rate (ASR), while stealth was quantified using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS).
4. Results and Analysis
4.1 Attack Effectiveness and Stealth
AS-FIBA achieved a high average ASR of 81.72% across the tested models. Despite this high success rate, the poisoned inputs remained visually indistinguishable from clean inputs. The PSNR of poisoned images remained around 21.98 dB, which the authors note shows minimal perceptual distortion in the context of these high-resolution restoration tasks.
4.2 Robustness to Defenses
The attack demonstrated significant resilience against common backdoor defense mechanisms:
- Fine-Pruning: While traditional attacks often fail when 70-80% of a model's neurons are pruned, AS-FIBA maintained a presence, suggesting the backdoor is distributed across the network's architecture.
- STRIP: The attack successfully bypassed STRIP (STRong Intentional Perturbation), a defense that identifies triggers by measuring the entropy of outputs when inputs are overlaid. Because AS-FIBA's triggers are adaptive and frequency-based, they do not produce the predictable entropy signatures that STRIP looks for.
5. Conclusion
AS-FIBA represents a new class of stealthy backdoor attacks tailored for the nuances of image restoration. By exploiting the frequency domain and targeting subtle quality degradation, the attack bypasses both human scrutiny and algorithmic defenses. These findings highlight a critical need for the development of frequency-aware defensive filters and more robust training protocols in the field of deep face restoration.
Summary updated