Traditional backdoor attacks, such as the seminal BadNets, rely on injecting a visible "trigger" (like a white square in the corner of an image) into training data. While effective, these attacks are easily thwarted by human inspection or automated anomaly detection. The research direction represented by these 136 papers shifts the focus toward stealthy and invisible backdoors, where the malicious trigger is indistinguishable from natural data or common processing artifacts.
The core intuition behind these attacks is that Deep Neural Networks (DNNs) are "shortcut learners." They will latch onto the simplest statistical correlation that minimizes loss. By embedding a trigger that is perceptible to the model's feature extractors but imperceptible to the human eye, an attacker can create a "hidden" decision boundary that only activates under specific, adversarial conditions.
Leveraging Natural Data Artifacts
One of the most potent ways to hide a backdoor is to repurpose existing, "natural" variations in data as triggers. Instead of adding an artificial patch, researchers like Yunfei Liu and Xingjun Ma demonstrate in Reflection Backdoor: A Natural Backdoor Attack on Deep Neural Networks that the physics of light—specifically reflections—can serve as a trigger. Because reflections occur naturally in real-world images, a model trained on poisoned "reflected" images learns to associate the presence of a reflection with a target class, bypassing human suspicion.
Similarly, Sze Jue Yang and Khoa D. Doan show that standard lossy compression (like JPEG or WebP) can be repurposed as a trigger in Everyone Can Attack: Repurpose Lossy Compression as a Natural Backdoor Attack. In this scenario, the "trigger" is simply the mathematical artifact of compression itself. In video domains, Hasan Abed Al Kader Hammoud and Bernard Ghanem exploit temporal artifacts like motion blur or frame lag to achieve the same effect in Look, Listen, and Attack: Backdoor Attacks Against Video Action Recognition.
Mathematical Invisibility via Frequency and Steganography
If a trigger cannot be a natural artifact, it must be mathematically hidden. Benjamin Zi Hao Zhao and Minhui Xue pioneered this using steganography and $L_p$-norm regularization in Invisible Backdoor Attacks Against Deep Neural Networks. By constraining the perturbation to be extremely small, the trigger becomes a form of high-frequency noise that the model can see but humans cannot.
Recent work has moved this into the frequency domain. Yi Yu and Wenhan Yang demonstrate in Backdoor Attacks Against Deep Image Compression via Adaptive Frequency Trigger that triggers can be embedded in the Discrete Cosine Transform (DCT) coefficients. Because humans are less sensitive to certain frequency bands, attackers can hide a potent signal in the "invisible" parts of the image spectrum. This is further refined by Honglong Chen in SFIBA: Spatial-based Full-target Invisible Backdoor Attacks, which uses Fast Fourier Transforms (FFT) and Singular Value Decomposition (SVD) to distribute the trigger across the entire image.
Structural and Contextual Deception
Beyond pixel-level changes, stealth can be achieved through structural transformations. Anh Tuan Nguyen and Anh Tuan Tran introduced WaNet in WaNet - Imperceptible Warping-based Backdoor Attack. WaNet uses subtle geometric warping (elastic transformations) as a trigger. Because the entire image is slightly distorted in a consistent way, there is no "patch" to find, and the distortion is often mistaken for natural camera lens effects.
In more complex tasks like semantic segmentation, Bilal Hussain Abbasi and Zirui Gong propose ConSeg in ConSeg: Contextual Backdoor Attack Against Semantic Segmentation. ConSeg exploits the context of a scene—for example, making a model misclassify a "pedestrian" only when they appear next to a specific, naturally occurring "target" object like a certain type of tree or mailbox.
Why Stealthy Backdoors Defeat Defenses
The shift toward invisibility creates a fundamental problem for AI safety. Most defenses, such as Anti-Backdoor Learning (ABL) by Yige Li and Lingjuan Lyu, rely on the fact that models learn simple, high-contrast triggers much faster than complex natural features.
However, as triggers become more "natural" or "invisible," this gap closes. When a trigger like motion blur or a reflection overlaps with the legitimate feature distribution of the clean data, the model no longer sees it as an anomaly. As Zhen Xiang and David J. Miller note in their work on Revealing Backdoors, if the "cost" of forcing a misclassification via a trigger is not unnaturally low compared to natural class variations, the backdoor becomes statistically indistinguishable from legitimate model behavior.