Greta Dolcetti, Giulio Zizzo, Sergio Maffeis
We present an experimental evaluation that assesses the robustness of four open source LLMs claiming function-calling capabilities against three different attacks, and we measure the effectiveness of eight different defences. Our results show how these models are not safe by default, and how the defences are not yet employable in real-world scenarios.
Ambrish Rawat, Stefan Schoepf, Giulio Zizzo et al.
As generative AI, particularly large language models (LLMs), become increasingly integrated into production applications, new attack surfaces and vulnerabilities emerge and put a focus on adversarial threats in natural language and multi-modal systems. Red-teaming has gained importance in proactively identifying weaknesses in these systems, while blue-teaming works to protect against such adversarial attacks. Despite growing academic interest in adversarial risks for generative AI, there is limited guidance tailored for practitioners to assess and mitigate these challenges in real-world environments. To address this, our contributions include: (1) a practical examination of red- and blue-teaming strategies for securing generative AI, (2) identification of key challenges and open questions in defense development and evaluation, and (3) the Attack Atlas, an intuitive framework that brings a practical approach to analyzing single-turn input attacks, placing it at the forefront for practitioners. This work aims to bridge the gap between academic insights and practical security measures for the protection of generative AI systems.
Yannis Belkhiter, Seshu Tirupathi, Giulio Zizzo et al.
The field of Language Reasoning Models (LRMs) has been very active over the past few years with advances in training and inference techniques enabling LRMs to reason longer, and more accurately. However, a growing body of studies show that LRMs are still inefficient, over-generating verification and reflection steps. To address this challenge, we introduce the Step-Tagging framework, a lightweight sentence-classifier enabling real-time annotation of the type of reasoning steps that an LRM is generating. To monitor reasoning behaviors, we introduced ReasonType: a novel taxonomy of reasoning steps. Building on this framework, we demonstrated that online monitoring of the count of specific steps can produce effective interpretable early stopping criteria of LRM inferences. We evaluate the Step-tagging framework on three open-source reasoning models across standard benchmark datasets: MATH500, GSM8K, AIME and non-mathematical tasks (GPQA and MMLU-Pro). We achieve 20 to 50\% token reduction while maintaining comparable accuracy to standard generation, with largest gains observed on more computation-heavy tasks. This work offers a novel way to increase control over the generation of LRMs, and a new tool to study behaviors of LRMs.
Myles Foley, Ambrish Rawat, Taesung Lee et al.
The wide applicability and adaptability of generative large language models (LLMs) has enabled their rapid adoption. While the pre-trained models can perform many tasks, such models are often fine-tuned to improve their performance on various downstream applications. However, this leads to issues over violation of model licenses, model theft, and copyright infringement. Moreover, recent advances show that generative technology is capable of producing harmful content which exacerbates the problems of accountability within model supply chains. Thus, we need a method to investigate how a model was trained or a piece of text was generated and what their pre-trained base model was. In this paper we take the first step to address this open problem by tracing back the origin of a given fine-tuned LLM to its corresponding pre-trained base model. We consider different knowledge levels and attribution strategies, and find that we can correctly trace back 8 out of the 10 fine tuned models with our best method.
Ambrish Rawat, Giulio Zizzo, Swanand Kadhe et al.
In this work, we devise robust and efficient learning protocols for orchestrating a Federated Learning (FL) process for the Federated Tumor Segmentation Challenge (FeTS 2022). Enabling FL for FeTS setup is challenging mainly due to data heterogeneity among collaborators and communication cost of training. To tackle these challenges, we propose Robust Learning Protocol (RoLePRO) which is a combination of server-side adaptive optimisation (e.g., server-side Adam) and judicious parameter (weights) aggregation schemes (e.g., adaptive weighted aggregation). RoLePRO takes a two-phase approach, where the first phase consists of vanilla Federated Averaging, while the second phase consists of a judicious aggregation scheme that uses a sophisticated reweighting, all in the presence of an adaptive optimisation algorithm at the server. We draw insights from extensive experimentation to tune learning rates for the two phases.
Giandomenico Cornacchia, Giulio Zizzo, Kieran Fraser et al.
The proliferation of Large Language Models (LLMs) in diverse applications underscores the pressing need for robust security measures to thwart potential jailbreak attacks. These attacks exploit vulnerabilities within LLMs, endanger data integrity and user privacy. Guardrails serve as crucial protective mechanisms against such threats, but existing models often fall short in terms of both detection accuracy, and computational efficiency. This paper advocates for the significance of jailbreak attack prevention on LLMs, and emphasises the role of input guardrails in safeguarding these models. We introduce MoJE (Mixture of Jailbreak Expert), a novel guardrail architecture designed to surpass current limitations in existing state-of-the-art guardrails. By employing simple linguistic statistical techniques, MoJE excels in detecting jailbreak attacks while maintaining minimal computational overhead during model inference. Through rigorous experimentation, MoJE demonstrates superior performance capable of detecting 90% of the attacks without compromising benign prompts, enhancing LLMs security against jailbreak attacks.
Daniel Gibert, Giulio Zizzo, Quan Le
Malware detectors based on deep learning (DL) have been shown to be susceptible to malware examples that have been deliberately manipulated in order to evade detection, a.k.a. adversarial malware examples. More specifically, it has been show that deep learning detectors are vulnerable to small changes on the input file. Given this vulnerability of deep learning detectors, we propose a practical defense against adversarial malware examples inspired by randomized smoothing. In our work, instead of employing Gaussian or Laplace noise when randomizing inputs, we propose a randomized ablation-based smoothing scheme that ablates a percentage of the bytes within an executable. During training, our randomized ablation-based smoothing scheme trains a base classifier based on ablated versions of the executable files. At test time, the final classification for a given input executable is taken as the class most commonly predicted by the classifier on a set of ablated versions of the original executable. To demonstrate the suitability of our approach we have empirically evaluated the proposed ablation-based model against various state-of-the-art evasion attacks on the BODMAS dataset. Results show greater robustness and generalization capabilities to adversarial malware examples in comparison to a non-smoothed classifier.
Yannis Belkhiter, Giulio Zizzo, Sergio Maffeis et al.
The growth of agentic AI has drawn significant attention to function calling Large Language Models (LLMs), which are designed to extend the capabilities of AI-powered system by invoking external functions. Injection and jailbreaking attacks have been extensively explored to showcase the vulnerabilities of LLMs to user prompt manipulation. The expanded capabilities of agentic models introduce further vulnerabilities via their function calling interface. Recent work in LLM security showed that function calling can be abused, leading to data tampering and theft, causing disruptive behavior such as endless loops, or causing LLMs to produce harmful content in the style of jailbreaking attacks. This paper introduces a novel function hijacking attack (FHA) that manipulates the tool selection process of agentic models to force the invocation of a specific, attacker-chosen function. While existing attacks focus on semantic preference of the model for function-calling tasks, we show that FHA is largely agnostic to the context semantics and robust to the function sets, making it applicable across diverse domains. We further demonstrate that FHA can be trained to produce universal adversarial functions, enabling a single attacked function to hijack tool selection across multiple queries and payload configurations. We conducted experiments on 5 different models, including instructed and reasoning variants, reaching 70% to 100% ASR over the established BFCL dataset. Our findings further demonstrate the need for strong guardrails and security modules for agentic systems.
Yannis Belkhiter, Seshu Tirupathi, Giulio Zizzo et al.
The field of Language Reasoning Models (LRMs) has been very active over the past few years with advances in training and inference techniques enabling LRMs to reason longer, and more accurately. However, a growing body of studies show that LRMs are still inefficient, over-generating verification and reflection steps. Additionally, the high-level role of each reasoning step and how different step types contribute to the generation of correct answers, is largely underexplored. To address this challenge, we introduce TRACES (Tagging of the Reasoning steps enabling Adaptive Cost-Efficient early-Stopping), a lightweight framework that tags reasoning steps in real-time, and enable adaptive, cost-efficient early stopping of large-language-model inferences. Building on this framework we monitor reasoning behaviors during inferences, and we find that LRMs tend to shift their reasoning behavior after reaching a correct answer. We demonstrate that the monitoring of the specific type of steps can produce effective interpretable early stopping criteria. We evaluate the TRACES framework on three mathematical reasoning benchmarks, namely, MATH500, GSM8K, AIME and two knowledge and reasoning benchmarks, MMLU and GPQA respectively. We achieve 20 to 50% token reduction while maintaining comparable accuracy to standard generation.
Inkit Padhi, Manish Nagireddy, Giandomenico Cornacchia et al. · ibm-research
We introduce the Granite Guardian models, a suite of safeguards designed to provide risk detection for prompts and responses, enabling safe and responsible use in combination with any large language model (LLM). These models offer comprehensive coverage across multiple risk dimensions, including social bias, profanity, violence, sexual content, unethical behavior, jailbreaking, and hallucination-related risks such as context relevance, groundedness, and answer relevance for retrieval-augmented generation (RAG). Trained on a unique dataset combining human annotations from diverse sources and synthetic data, Granite Guardian models address risks typically overlooked by traditional risk detection models, such as jailbreaks and RAG-specific issues. With AUC scores of 0.871 and 0.854 on harmful content and RAG-hallucination-related benchmarks respectively, Granite Guardian is the most generalizable and competitive model available in the space. Released as open-source, Granite Guardian aims to promote responsible AI development across the community. https://github.com/ibm-granite/granite-guardian
Giulio Zizzo, Giandomenico Cornacchia, Kieran Fraser et al.
As large language models (LLMs) become integrated into everyday applications, ensuring their robustness and security is increasingly critical. In particular, LLMs can be manipulated into unsafe behaviour by prompts known as jailbreaks. The variety of jailbreak styles is growing, necessitating the use of external defences known as guardrails. While many jailbreak defences have been proposed, not all defences are able to handle new out-of-distribution attacks due to the narrow segment of jailbreaks used to align them. Moreover, the lack of systematisation around defences has created significant gaps in their practical application. In this work, we perform systematic benchmarking across 15 different defences, considering a broad swathe of malicious and benign datasets. We find that there is significant performance variation depending on the style of jailbreak a defence is subject to. Additionally, we show that based on current datasets available for evaluation, simple baselines can display competitive out-of-distribution performance compared to many state-of-the-art defences. Code is available at https://github.com/IBM/Adversarial-Prompt-Evaluation.
Kristjan Greenewald, Luis Lastras, Thomas Parnell et al.
Low-Rank Adaptation (LoRA) has emerged as a highly efficient framework for finetuning the weights of large foundation models, and has become the go-to method for data-driven customization of LLMs. Despite the promise of highly customized behaviors and capabilities, switching between relevant LoRAs in a multiturn setting is inefficient, as the key-value (KV) cache of the entire turn history must be recomputed with the LoRA weights before generation can begin. To address this problem, we propose Activated LoRA (aLoRA), an adapter architecture which modifies the LoRA framework to only adapt weights for the tokens in the sequence after the aLoRA is invoked. This change crucially allows aLoRA to accept the base model's KV cache of the input string, meaning that aLoRA can be instantly activated whenever needed in a chain without recomputing the prior keys and values. This enables building what we call intrinsics, i.e. specialized models invoked to perform well-defined operations on portions of an input chain or conversation that otherwise uses the base model by default. We train a set of aLoRA-based intrinsics models, demonstrating competitive accuracy with standard LoRA while significantly improving inference efficiency. We contributed our Activated LoRA implementation to the Huggingface PEFT library https://github.com/huggingface/peft.
Subina Khanal, Seshu Tirupathi, Giulio Zizzo et al.
The recent breakthrough of Transformers in deep learning has drawn significant attention of the time series community due to their ability to capture long-range dependencies. However, like other deep learning models, Transformers face limitations in time series prediction, including insufficient temporal understanding, generalization challenges, and data shift issues for the domains with limited data. Additionally, addressing the issue of catastrophic forgetting, where models forget previously learned information when exposed to new data, is another critical aspect that requires attention in enhancing the robustness of Transformers for time series tasks. To address these limitations, in this paper, we pre-train the time series Transformer model on a source domain with sufficient data and fine-tune it on the target domain with limited data. We introduce the \emph{One-step fine-tuning} approach, adding some percentage of source domain data to the target domains, providing the model with diverse time series instances. We then fine-tune the pre-trained model using a gradual unfreezing technique. This helps enhance the model's performance in time series prediction for domains with limited data. Extensive experimental results on two real-world datasets show that our approach improves over the state-of-the-art baselines by 4.35% and 11.54% for indoor temperature and wind power prediction, respectively.
Daniel Gibert, Giulio Zizzo, Quan Le et al.
Deep learning-based malware detectors have been shown to be susceptible to adversarial malware examples, i.e. malware examples that have been deliberately manipulated in order to avoid detection. In light of the vulnerability of deep learning detectors to subtle input file modifications, we propose a practical defense against adversarial malware examples inspired by (de)randomized smoothing. In this work, we reduce the chances of sampling adversarial content injected by malware authors by selecting correlated subsets of bytes, rather than using Gaussian noise to randomize inputs like in the Computer Vision (CV) domain. During training, our ablation-based smoothing scheme trains a base classifier to make classifications on a subset of contiguous bytes or chunk of bytes. At test time, a large number of chunks are then classified by a base classifier and the consensus among these classifications is then reported as the final prediction. We propose two strategies to determine the location of the chunks used for classification: (1) randomly selecting the locations of the chunks and (2) selecting contiguous adjacent chunks. To showcase the effectiveness of our approach, we have trained two classifiers with our chunk-based ablation schemes on the BODMAS dataset. Our findings reveal that the chunk-based smoothing classifiers exhibit greater resilience against adversarial malware examples generated with state-of-the-are evasion attacks, outperforming a non-smoothed classifier and a randomized smoothing-based classifier by a great margin.
Daniel Gibert, Luca Demetrio, Giulio Zizzo et al.
Deep learning-based malware detection systems are vulnerable to adversarial EXEmples - carefully-crafted malicious programs that evade detection with minimal perturbation. As such, the community is dedicating effort to develop mechanisms to defend against adversarial EXEmples. However, current randomized smoothing-based defenses are still vulnerable to attacks that inject blocks of adversarial content. In this paper, we introduce a certifiable defense against patch attacks that guarantees, for a given executable and an adversarial patch size, no adversarial EXEmple exist. Our method is inspired by (de)randomized smoothing which provides deterministic robustness certificates. During training, a base classifier is trained using subsets of continguous bytes. At inference time, our defense splits the executable into non-overlapping chunks, classifies each chunk independently, and computes the final prediction through majority voting to minimize the influence of injected content. Furthermore, we introduce a preprocessing step that fixes the size of the sections and headers to a multiple of the chunk size. As a consequence, the injected content is confined to an integer number of chunks without tampering the other chunks containing the real bytes of the input examples, allowing us to extend our certified robustness guarantees to content insertion attacks. We perform an extensive ablation study, by comparing our defense with randomized smoothing-based defenses against a plethora of content manipulation attacks and neural network architectures. Results show that our method exhibits unmatched robustness against strong content-insertion attacks, outperforming randomized smoothing-based defenses in the literature.
Yannis Belkhiter, Giulio Zizzo, Sergio Maffeis
With the introduction of the transformers architecture, LLMs have revolutionized the NLP field with ever more powerful models. Nevertheless, their development came up with several challenges. The exponential growth in computational power and reasoning capabilities of language models has heightened concerns about their security. As models become more powerful, ensuring their safety has become a crucial focus in research. This paper aims to address gaps in the current literature on jailbreaking techniques and the evaluation of LLM vulnerabilities. Our contributions include the creation of a novel dataset designed to assess the harmfulness of model outputs across multiple harm levels, as well as a focus on fine-grained harm-level analysis. Using this framework, we provide a comprehensive benchmark of state-of-the-art jailbreaking attacks, specifically targeting the Vicuna 13B v1.5 model. Additionally, we examine how quantization techniques, such as AWQ and GPTQ, influence the alignment and robustness of models, revealing trade-offs between enhanced robustness with regards to transfer attacks and potential increases in vulnerability on direct ones. This study aims to demonstrate the influence of harmful input queries on the complexity of jailbreaking techniques, as well as to deepen our understanding of LLM vulnerabilities and improve methods for assessing model robustness when confronted with harmful content, particularly in the context of compression strategies.
Janvi Thakkar, Giulio Zizzo, Sergio Maffeis
Machine learning models are being used in an increasing number of critical applications; thus, securing their integrity and ownership is critical. Recent studies observed that adversarial training and watermarking have a conflicting interaction. This work introduces a novel framework to integrate adversarial training with watermarking techniques to fortify against evasion attacks and provide confident model verification in case of intellectual property theft. We use adversarial training together with adversarial watermarks to train a robust watermarked model. The key intuition is to use a higher perturbation budget to generate adversarial watermarks compared to the budget used for adversarial training, thus avoiding conflict. We use the MNIST and Fashion-MNIST datasets to evaluate our proposed technique on various model stealing attacks. The results obtained consistently outperform the existing baseline in terms of robustness performance and further prove the resilience of this defense against pruning and fine-tuning removal attacks.
Daniel Gibert, Nikolaos Totosis, Constantinos Patsakis et al.
The proliferation of malware, particularly through the use of packing, presents a significant challenge to static analysis and signature-based malware detection techniques. The application of packing to the original executable code renders extracting meaningful features and signatures challenging. To deal with the increasing amount of malware in the wild, researchers and anti-malware companies started harnessing machine learning capabilities with very promising results. However, little is known about the effects of packing on static machine learning-based malware detection and classification systems. This work addresses this gap by investigating the impact of packing on the performance of static machine learning-based models used for malware detection and classification, with a particular focus on those using visualisation techniques. To this end, we present a comprehensive analysis of various packing techniques and their effects on the performance of machine learning-based detectors and classifiers. Our findings highlight the limitations of current static detection and classification systems and underscore the need to be proactive to effectively counteract the evolving tactics of malware authors.
Yannis Belkhiter, Seshu Tirupathi, Giulio Zizzo et al.
As AI becomes a native component of 6G network control, AI models must adapt to continuously changing conditions, including the introduction of new features and measurements driven by multi-vendor deployments, hardware upgrades, and evolving service requirements. To address this growing need for flexible learning in non-stationary environments, this vision paper highlights Adaptive Random Forests (ARFs) as a reliable solution for dynamic feature adaptation in communication network scenarios. We show that iterative training of ARFs can effectively lead to stable predictions, with accuracy improving over time as more features are added. In addition, we highlight the importance of explainability in AI-driven networks, proposing Drift-Aware Feature Importance (DAFI) as an efficient XAI feature importance (FI) method. DAFI uses a distributional drift detector to signal when to apply computationally intensive FI methods instead of lighter alternatives. Our tests on 3 different datasets indicate that our approach reduces runtime by up to 2 times, while producing more consistent feature importance values. Together, ARFs and DAFI provide a promising framework to build flexible AI methods adapted to 6G network use-cases.
Yannis Belkhiter, Seshu Tirupathi, Giulio Zizzo et al.
The field of AutoML has made remarkable progress in post-hoc model selection, with libraries capable of automatically identifying the most performing models for a given dataset. Nevertheless, these methods often rely on exhaustive hyperparameter searches, where methods automatically train and test different types of models on the target dataset. Contrastingly, pre-hoc prediction emerges as a promising alternative, capable of bypassing exhaustive search through intelligent pre-selection of models. Despite its potential, pre-hoc prediction remains under-explored in the literature. This paper explores the intersection of AutoML and pre-hoc model selection by leveraging traditional models and Large Language Model (LLM) agents to reduce the search space of AutoML libraries. By relying on dataset descriptions and statistical information, we reduce the AutoML search space. Our methodology is applied to the AWS AutoGluon portfolio dataset, a state-of-the-art AutoML benchmark containing 175 tabular classification datasets available on OpenML. The proposed approach offers a shift in AutoML workflows, significantly reducing computational overhead, while still selecting the best model for the given dataset.
Stefan Schoepf, Muhammad Zaid Hameed, Ambrish Rawat et al.
With LLM usage rapidly increasing, their vulnerability to jailbreaks that create harmful outputs are a major security risk. As new jailbreaking strategies emerge and models are changed by fine-tuning, continuous testing for security vulnerabilities is necessary. Existing Red Teaming methods fall short in cost efficiency, attack success rate, attack diversity, or extensibility as new attack types emerge. We address these challenges with Modular And Diverse Malicious Attack MiXtures (MAD-MAX) for Automated LLM Red Teaming. MAD-MAX uses automatic assignment of attack strategies into relevant attack clusters, chooses the most relevant clusters for a malicious goal, and then combines strategies from the selected clusters to achieve diverse novel attacks with high attack success rates. MAD-MAX further merges promising attacks together at each iteration of Red Teaming to boost performance and introduces a similarity filter to prune out similar attacks for increased cost efficiency. The MAD-MAX approach is designed to be easily extensible with newly discovered attack strategies and outperforms the prominent Red Teaming method Tree of Attacks with Pruning (TAP) significantly in terms of Attack Success Rate (ASR) and queries needed to achieve jailbreaks. MAD-MAX jailbreaks 97% of malicious goals in our benchmarks on GPT-4o and Gemini-Pro compared to TAP with 66%. MAD-MAX does so with only 10.9 average queries to the target LLM compared to TAP with 23.3. WARNING: This paper contains contents which are offensive in nature.
Janvi Thakkar, Giulio Zizzo, Sergio Maffeis
Malicious adversaries can attack machine learning models to infer sensitive information or damage the system by launching a series of evasion attacks. Although various work addresses privacy and security concerns, they focus on individual defenses, but in practice, models may undergo simultaneous attacks. This study explores the combination of adversarial training and differentially private training to defend against simultaneous attacks. While differentially-private adversarial training, as presented in DP-Adv, outperforms the other state-of-the-art methods in performance, it lacks formal privacy guarantees and empirical validation. Thus, in this work, we benchmark the performance of this technique using a membership inference attack and empirically show that the resulting approach is as private as non-robust private models. This work also highlights the need to explore privacy guarantees in dynamic training paradigms.
Giulio Zizzo, Ambrish Rawat, Mathieu Sinn et al.
In federated learning (FL), robust aggregation schemes have been developed to protect against malicious clients. Many robust aggregation schemes rely on certain numbers of benign clients being present in a quorum of workers. This can be hard to guarantee when clients can join at will, or join based on factors such as idle system status, and connected to power and WiFi. We tackle the scenario of securing FL systems conducting adversarial training when a quorum of workers could be completely malicious. We model an attacker who poisons the model to insert a weakness into the adversarial training such that the model displays apparent adversarial robustness, while the attacker can exploit the inserted weakness to bypass the adversarial training and force the model to misclassify adversarial examples. We use abstract interpretation techniques to detect such stealthy attacks and block the corrupted model updates. We show that this defence can preserve adversarial robustness even against an adaptive attacker.
Giulio Zizzo, Ambrish Rawat, Mathieu Sinn et al.
Federated learning (FL) is one of the most important paradigms addressing privacy and data governance issues in machine learning (ML). Adversarial training has emerged, so far, as the most promising approach against evasion threats on ML models. In this paper, we take the first known steps towards federated adversarial training (FAT) combining both methods to reduce the threat of evasion during inference while preserving the data privacy during training. We investigate the effectiveness of the FAT protocol for idealised federated settings using MNIST, Fashion-MNIST, and CIFAR10, and provide first insights on stabilising the training on the LEAF benchmark dataset which specifically emulates a federated learning environment. We identify challenges with this natural extension of adversarial training with regards to achieved adversarial robustness and further examine the idealised settings in the presence of clients undermining model convergence. We find that Trimmed Mean and Bulyan defences can be compromised and we were able to subvert Krum with a novel distillation based attack which presents an apparently "robust" model to the defender while in fact the model fails to provide robustness against simple attack modifications.
Giulio Zizzo, Chris Hankin, Sergio Maffeis et al.
Neural networks are increasingly used for intrusion detection on industrial control systems (ICS). With neural networks being vulnerable to adversarial examples, attackers who wish to cause damage to an ICS can attempt to hide their attacks from detection by using adversarial example techniques. In this work we address the domain specific challenges of constructing such attacks against autoregressive based intrusion detection systems (IDS) in an ICS setting. We model an attacker that can compromise a subset of sensors in a ICS which has a LSTM based IDS. The attacker manipulates the data sent to the IDS, and seeks to hide the presence of real cyber-physical attacks occurring in the ICS. We evaluate our adversarial attack methodology on the Secure Water Treatment system when examining solely continuous data, and on data containing a mixture of discrete and continuous variables. In the continuous data domain our attack successfully hides the cyber-physical attacks requiring 2.87 out of 12 monitored sensors to be compromised on average. With both discrete and continuous data our attack required, on average, 3.74 out of 26 monitored sensors to be compromised.
Giulio Zizzo, Chris Hankin, Sergio Maffeis et al.
Deep learning methods have shown state of the art performance in a range of tasks from computer vision to natural language processing. However, it is well known that such systems are vulnerable to attackers who craft inputs in order to cause misclassification. The level of perturbation an attacker needs to introduce in order to cause such a misclassification can be extremely small, and often imperceptible. This is of significant security concern, particularly where misclassification can cause harm to humans. We thus propose Deep Latent Defence, an architecture which seeks to combine adversarial training with a detection system. At its core Deep Latent Defence has a adversarially trained neural network. A series of encoders take the intermediate layer representation of data as it passes though the network and project it to a latent space which we use for detecting adversarial samples via a $k$-nn classifier. We present results using both grey and white box attackers, as well as an adaptive $L_{\infty}$ bounded attack which was constructed specifically to try and evade our defence. We find that even under the strongest attacker model that we have investigated our defence is able to offer significant defensive benefits.