Most of the approaches proposed so far to craft targeted adversarial examples against Deep Learning classifiers are highly suboptimal and typically rely on increasing the likelihood of the target class, thus implicitly focusing on one-hot encoding settings. In this paper, we propose a more general, theoretically sound, targeted attack that resorts to the minimization of a Jacobian-induced MAhalanobis distance (JMA) term, taking into account the effort (in the input space) required to move the latent space representation of the input sample in a given direction. The minimization is solved by exploiting the Wolfe duality theorem, reducing the problem to the solution of a Non-Negative Least Square (NNLS) problem. The proposed algorithm provides an optimal solution to a linearized version of the adversarial example problem originally introduced by Szegedy et al. \cite{szegedy2013intriguing}. The experiments we carried out confirm the generality of the proposed attack which is proven to be effective under a wide variety of output encoding schemes. Noticeably, the JMA attack is also effective in a multi-label classification scenario, being capable to induce a targeted modification of up to half the labels in a complex multilabel classification scenario with 20 labels, a capability that is out of reach of all the attacks proposed so far. As a further advantage, the JMA attack usually requires very few iterations, thus resulting more efficient than existing methods.
In recent years, there has been significant growth in the commercial applications of generative models, licensed and distributed by model developers to users, who in turn use them to offer services. In this scenario, there is a need to track and identify the responsible user in the presence of a violation of the license agreement or any kind of malicious usage. Although there are methods enabling Generative Adversarial Networks (GANs) to include invisible watermarks in the images they produce, generating a model with a different watermark, referred to as a fingerprint, for each user is time- and resource-consuming due to the need to retrain the model to include the desired fingerprint. In this paper, we propose a retraining-free GAN fingerprinting method that allows model developers to easily generate model copies with the same functionality but different fingerprints. The generator is modified by inserting additional Personalized Normalization (PN) layers whose parameters (scaling and bias) are generated by two dedicated shallow networks (ParamGen Nets) taking the fingerprint as input. A watermark decoder is trained simultaneously to extract the fingerprint from the generated images. The proposed method can embed different fingerprints inside the GAN by just changing the input of the ParamGen Nets and performing a feedforward pass, without finetuning or retraining. The performance of the proposed method in terms of robustness against both model-level and image-level attacks is also superior to the state-of-the-art.
We propose a novel multi-bit box-free watermarking method for the protection of Intellectual Property Rights (IPR) of GANs with improved robustness against white-box attacks like fine-tuning, pruning, quantization, and surrogate model attacks. The watermark is embedded by adding an extra watermarking loss term during GAN training, ensuring that the images generated by the GAN contain an invisible watermark that can be retrieved by a pre-trained watermark decoder. In order to improve the robustness against white-box model-level attacks, we make sure that the model converges to a wide flat minimum of the watermarking loss term, in such a way that any modification of the model parameters does not erase the watermark. To do so, we add random noise vectors to the parameters of the generator and require that the watermarking loss term is as invariant as possible with respect to the presence of noise. This procedure forces the generator to converge to a wide flat minimum of the watermarking loss. The proposed method is architectureand dataset-agnostic, thus being applicable to many different generation tasks and models, as well as to CNN-based image processing architectures. We present the results of extensive experiments showing that the presence of the watermark has a negligible impact on the quality of the generated images, and proving the superior robustness of the watermark against model modification and surrogate model attacks.
Despite the wide variety of methods developed for synthetic image attribution, most of them can only attribute images generated by models or architectures included in the training set and do not work with unknown architectures, hindering their applicability in real-world scenarios. In this paper, we propose a verification framework that relies on a Siamese Network to address the problem of open-set attribution of synthetic images to the architecture that generated them. We consider two different settings. In the first setting, the system determines whether two images have been produced by the same generative architecture or not. In the second setting, the system verifies a claim about the architecture used to generate a synthetic image, utilizing one or multiple reference images generated by the claimed architecture. The main strength of the proposed system is its ability to operate in both closed and open-set scenarios so that the input images, either the query and reference images, can belong to the architectures considered during training or not. Experimental evaluations encompassing various generative architectures such as GANs, diffusion models, and transformers, focusing on synthetic face image generation, confirm the excellent performance of our method in both closed and open-set settings, as well as its strong generalization capabilities.
Classification of AI-manipulated content is receiving great attention, for distinguishing different types of manipulations. Most of the methods developed so far fail in the open-set scenario, that is when the algorithm used for the manipulation is not represented by the training set. In this paper, we focus on the classification of synthetic face generation and manipulation in open-set scenarios, and propose a method for classification with a rejection option. The proposed method combines the use of Vision Transformers (ViT) with a hybrid approach for simultaneous classification and localization. Feature map correlation is exploited by the ViT module, while a localization branch is employed as an attention mechanism to force the model to learn per-class discriminative features associated with the forgery when the manipulation is performed locally in the image. Rejection is performed by considering several strategies and analyzing the model output layers. The effectiveness of the proposed method is assessed for the task of classification of facial attribute editing and GAN attribution.
In this paper, we propose a Universal Defence based on Clustering and Centroids Analysis (CCA-UD) against backdoor attacks. The goal of the proposed defence is to reveal whether a Deep Neural Network model is subject to a backdoor attack by inspecting the training dataset. CCA-UD first clusters the samples of the training set by means of density-based clustering. Then, it applies a novel strategy to detect the presence of poisoned clusters. The proposed strategy is based on a general misclassification behaviour obtained when the features of a representative example of the analysed cluster are added to benign samples. The capability of inducing a misclassification error is a general characteristic of poisoned samples, hence the proposed defence is attack-agnostic. This mask a significant difference with respect to existing defences, that, either can defend against only some types of backdoor attacks, e.g., when the attacker corrupts the label of the poisoned samples, or are effective only when some conditions on the poisoning ratios adopted by the attacker or the kind of triggering pattern used by the attacker are satisfied. Experiments carried out on several classification tasks, considering different types of backdoor attacks and triggering patterns, including both local and global triggers, reveal that the proposed method is very effective to defend against backdoor attacks in all the cases, always outperforming the state of the art techniques.
Due to the reduction of technological costs and the increase of satellites launches, satellite images are becoming more popular and easier to obtain. Besides serving benevolent purposes, satellite data can also be used for malicious reasons such as misinformation. As a matter of fact, satellite images can be easily manipulated relying on general image editing tools. Moreover, with the surge of Deep Neural Networks (DNNs) that can generate realistic synthetic imagery belonging to various domains, additional threats related to the diffusion of synthetically generated satellite images are emerging. In this paper, we review the State of the Art (SOTA) on the generation and manipulation of satellite images. In particular, we focus on both the generation of synthetic satellite imagery from scratch, and the semantic manipulation of satellite images by means of image-transfer technologies, including the transformation of images obtained from one type of sensor to another one. We also describe forensic detection techniques that have been researched so far to classify and detect synthetic image forgeries. While we focus mostly on forensic techniques explicitly tailored to the detection of AI-generated synthetic contents, we also review some methods designed for general splicing detection, which can in principle also be used to spot AI manipulate images
We propose a watermarking method for protecting the Intellectual Property (IP) of Generative Adversarial Networks (GANs). The aim is to watermark the GAN model so that any image generated by the GAN contains an invisible watermark (signature), whose presence inside the image can be checked at a later stage for ownership verification. To achieve this goal, a pre-trained CNN watermarking decoding block is inserted at the output of the generator. The generator loss is then modified by including a watermark loss term, to ensure that the prescribed watermark can be extracted from the generated images. The watermark is embedded via fine-tuning, with reduced time complexity. Results show that our method can effectively embed an invisible watermark inside the generated images. Moreover, our method is a general one and can work with different GAN architectures, different tasks, and different resolutions of the output image. We also demonstrate the good robustness performance of the embedded watermark against several post-processing, among them, JPEG compression, noise addition, blurring, and color transformations.
Predicting the country where a picture has been taken from has many potential applications, like detection of false claims, impostors identification, prevention of disinformation campaigns, identification of fake news and so on. Previous works have focused mostly on the estimation of the geo-coordinates where a picture has been taken. Yet, recognizing the country where an image has been taken could potentially be more important, from a semantic and forensic point of view, than identifying its spatial coordinates. So far only a few works have addressed this task, mostly by relying on images containing characteristic landmarks, like iconic monuments. In the above framework, this paper provides two main contributions. First, we introduce a new dataset, the VIPPGeo dataset, containing almost 4 million images, that can be used to train DL models for country classification. The dataset contains only urban images given the relevance of this kind of image for country recognition, and it has been built by paying attention to removing non-significant images, like images portraying faces or specific, non-relevant objects, like airplanes or ships. Secondly, we used the dataset to train a deep learning architecture casting the country recognition problem as a classification problem. The experiments, we performed, show that our network provides significantly better results than current state of the art. In particular, we found that asking the network to directly identify the country provides better results than estimating the geo-coordinates first and then using them to trace back to the country where the picture was taken.
Watermarking has been proposed as a way to protect the Intellectual Property Rights (IPR) of Deep Neural Networks (DNNs) and track their use. Several methods have been proposed that embed the watermark into the trainable parameters of the network (white box watermarking) or into the input-output mappping implemented by the network in correspondence to specific inputs (black box watermarking). In both cases, achieving robustness against fine tuning, model compression and, even more, transfer learning, is one of the most difficult challenges researchers are trying to face with. In this paper, we propose a new white-box, multi-bit watermarking algorithm with strong robustness properties, including retraining for transfer learning. Robustness is achieved thanks to a new information coding strategy according to which the watermark message is spread across a number of fixed weights, whose position depends on a secret key. The weights hosting the watermark are set prior to training, and are left unchanged throughout the entire training procedure. The distribution of the weights carrying out the message is theoretically optimised to make sure that the watermarked weights are indistinguishable from the other weights, while at the same time keeping their amplitude as large as possible to improve robustness against retraining. We carried out several experiments demonstrating the capability of the proposed scheme to provide high payloads with practically no impact on the network accuracy, at the same time retaining excellent robustness against network modifications an re-use, including retraining for transfer learning.