Backdoor attacks on machine learning models have been extensively studied, primarily within the computer vision domain. Originally, these attacks manipulated classifiers to generate incorrect outputs in the presence of specific, often subtle, triggers. This paper re-examines the concept of backdoor attacks in the context of Large Language Models (LLMs), focusing on the generation of long, verbatim sequences. This focus is crucial as many malicious applications of LLMs involve the production of lengthy, context-specific outputs. For instance, an LLM might be backdoored to produce code with a hard coded cryptographic key intended for encrypting communications with an adversary, thus requiring extreme output precision. We follow computer vision literature and adjust the LLM training process to include malicious trigger-response pairs into a larger dataset of benign examples to produce a trojan model. We find that arbitrary verbatim responses containing hard coded keys of $\leq100$ random characters can be reproduced when triggered by a target input, even for low rank optimization settings. Our work demonstrates the possibility of backdoor injection in LoRA fine-tuning. Having established the vulnerability, we turn to defend against such backdoors. We perform experiments on Gemini Nano 1.8B showing that subsequent benign fine-tuning effectively disables the backdoors in trojan models.

Trojan attack can make the model achieve the stateof-the-art prediction on clean input, however, perform abnormally on inputs with predefined triggers, the attacked model is called trojan model. Fig 1 shows the trojan attack examples: if you only input the black font sentence (clean input), the trojan model will output the normal prediction label, modifies Layer-wise Relevance Propagation and while you insert the specific trigger (red font) to head confidence to indicate head importance on sentence, the trojan model will output the flipped translation task, but it's not the case on many other label.

Due to data's unavailability or large size, and the high computational and human labor costs of training machine learning models, it is a common practice to rely on open source pre-trained models whenever possible. However, this practice is worry some from the security perspective. Pre-trained models can be infected with Trojan attacks, in which the attacker embeds a trigger in the model such that the model's behavior can be controlled by the attacker when the trigger is present in the input. In this paper, we present our preliminary work on a novel method for Trojan model detection. Our method creates a signature for a model based on activation optimization. A classifier is then trained to detect a Trojan model given its signature. Our method achieves state of the art performance on two public datasets.

Deep Neural Networks (DNN) are becoming increasingly more important in assisted and automated driving. Using such entities which are obtained using machine learning is inevitable: tasks such as recognizing traffic signs cannot be developed reasonably using traditional software development methods. DNN however do have the problem that they are mostly black boxes and therefore hard to understand and debug. One particular problem is that they are prone to hidden backdoors. This means that the DNN misclassifies its input, because it considers properties that should not be decisive for the output. Backdoors may either be introduced by malicious attackers or by inappropriate training. In any case, detecting and removing them is important in the automotive area, as they might lead to safety violations with potentially severe consequences. In this paper, we introduce a novel method to remove backdoors. Our method works for both intentional as well as unintentional backdoors. We also do not require prior knowledge about the shape or distribution of backdoors. Experimental evidence shows that our method performs well on several medium-sized examples.

Neural backdoors represent one primary threat to the security of deep learning systems. The intensive research has produced a plethora of backdoor attacks/defenses, resulting in a constant arms race. However, due to the lack of evaluation benchmarks, many critical questions remain under-explored: (i) what are the strengths and limitations of different attacks/defenses? (ii) what are the best practices to operate them? and (iii) how can the existing attacks/defenses be further improved? To bridge this gap, we design and implement TROJANZOO, the first open-source platform for evaluating neural backdoor attacks/defenses in a unified, holistic, and practical manner. Thus far, focusing on the computer vision domain, it has incorporated 8 representative attacks, 14 state-of-the-art defenses, 6 attack performance metrics, 10 defense utility metrics, as well as rich tools for in-depth analysis of the attack-defense interactions. Leveraging TROJANZOO, we conduct a systematic study on the existing attacks/defenses, unveiling their complex design spectrum: both manifest intricate trade-offs among multiple desiderata (e.g., the effectiveness, evasiveness, and transferability of attacks). We further explore improving the existing attacks/defenses, leading to a number of interesting findings: (i) one-pixel triggers often suffice; (ii) training from scratch often outperforms perturbing benign models to craft trojan models; (iii) optimizing triggers and trojan models jointly greatly improves both attack effectiveness and evasiveness; (iv) individual defenses can often be evaded by adaptive attacks; and (v) exploiting model interpretability significantly improves defense robustness. We envision that TROJANZOO will serve as a valuable platform to facilitate future research on neural backdoors.

Machine learning models in the wild have been shown to be vulnerable to Trojan attacks during training. Although many detection mechanisms have been proposed, strong adaptive attackers have been shown to be effective against them. In this paper, we aim to answer the questions considering an intelligent and adaptive adversary: (i) What is the minimal amount of instances required to be Trojaned by a strong attacker? and (ii) Is it possible for such an attacker to bypass strong detection mechanisms? We provide an analytical characterization of adversarial capability and strategic interactions between the adversary and detection mechanism that take place in such models. We characterize adversary capability in terms of the fraction of the input dataset that can be embedded with a Trojan trigger. We show that the loss function has a submodular structure, which leads to the design of computationally efficient algorithms to determine this fraction with provable bounds on optimality. We propose a Submodular Trojan algorithm to determine the minimal fraction of samples to inject a Trojan trigger. To evade detection of the Trojaned model, we model strategic interactions between the adversary and Trojan detection mechanism as a two-player game. We show that the adversary wins the game with probability one, thus bypassing detection. We establish this by proving that output probability distributions of a Trojan model and a clean model are identical when following the Min-Max (MM) Trojan algorithm. We perform extensive evaluations of our algorithms on MNIST, CIFAR-10, and EuroSAT datasets. The results show that (i) with Submodular Trojan algorithm, the adversary needs to embed a Trojan trigger into a very small fraction of samples to achieve high accuracy on both Trojan and clean samples, and (ii) the MM Trojan algorithm yields a trained Trojan model that evades detection with probability 1.