As large vision-language models (VLMs) see increasing adoption, the need for robust ownership verification becomes essential. However, training or fine-tuning these models is computationally intensive, making traditional watermarking approaches impractical. To address this, we introduce a lightweight, black-box watermarking technique that focuses solely on the vision encoder of the VLM, significantly reducing the resource requirements. By leveraging the bi-modal capabilities of VLMs, our method embeds owner-specific information by pairing a particular text prompt with a relevant image. When presented with this unique text-image combination and asked, "Who is your owner?", the model produces an explicit answer identifying its owner. This does not require access to model weights. We validated our approach on several models and demonstrated how it enables efficient and reliable watermark verification, all while bypassing the need to retrain or fine-tune the entire model.

International audience

Machine Learning is a well established tool used in a variety of applications. As training advanced models requires considerable amounts of meaningful data in addition to specific knowledge, a new business model separate models creators from model users. Pre-trained models are sold or made available as a service. This raises several security challenges, among others the one of intellectual property protection. Therefore, a new research track actively seeks to provide techniques for model watermarking that would enable model identification in case of suspicion of model theft or misuse. In this paper, we focus on the problem of secure watermarks verification, which affects all of the proposed techniques and until now was barely tackled. First, we revisit the existing threat model. In particular, we explain the possible threats related to a semi-honest or dishonest verification authority. Secondly, we show how to reduce trust requirements between participants by performing the watermarks verification on encrypted data. Finally, we describe a novel secure verification protocol as well as detail its possible implementation using Multi-Party Computation. The proposed solution does not only preserve the confidentiality of the watermarks but also helps detecting evasion attacks. It could be adopted to work with other authentication schemes based on watermarking, especially with image watermarking schemes.

In order to supervise the security of a large infrastructure, the administrator deploys multiple sensors and intrusion detection systems on several critical places in the system. It is easier to explain and detect attacks if more events are logged. Starting from a suspicious event (appearing as a log entry), the administrator can start his investigation by manually building the set of previous events that are linked to this event of interest. Accordingly, the administrator attempts to identify links among the logged events in order to retrieve those that correspond to the traces of the attacker's actions in the supervised system; previous work is aimed at building these connections. In practice, however, this type of link is not trivial to define and discover. Hence, there is a real necessity to describe and define formally the semantics of these links in literature. In this paper, a clear definition of this relationship, called contextual event causal dependency, is introduced and proposed. The work presented in this paper aims at defining a formal model that would ideally unify previous work on causal dependencies among heterogeneous events. We define a relationship among events that enables the discovery of all events, which can be considered as the cause (in the past) or the effect (in the future) of an event of interest(e.g., an indicator of compromise, produced by an attacker action). This model is gradually introduced and defined by merging two previously defined causality models from the distributed system and operating system research areas (i.e., Lamport's and d'Ausbourg's). Our model takes into consideration heterogeneous events that emanate from different abstraction layers (e.g., network, system, and application) with the main objective of formally defining a causal relationship among logged events. Thereafter, we show how existing implementations separately allow the computation of parts of the model. Finally, we describe the implementation and assessment of the model according to real attacks on distributed environments and its accuracy to extract all causally linked events related to a given attack event trace.

Nowadays, most of Web services are accessed through HTTPS. While preserving user privacy is important, it is also mandatory to monitor and detect specific users' actions, for instance, according to a security policy. This paper presents a solution to monitor HTTP/2 traffic over TLS. It highly differs from HTTP/1.1 over TLS traffic what makes existing monitoring techniques obsolete. Our solution, H2Classifier, aims at detecting if a user performs an action that has been previously defined over a monitored Web service, but without using any decryption. It is thus only based on passive traffic analysis and relies on random forest classifier. A challenge is to extract representative values of the loaded content associated to a Web page, which is actually customized based on the user action. Extensive evaluations with five top used Web services demonstrate the viability of our technique with an accuracy between 94% and 99%.

Network operators are currently very cautious before deploying a new network equipment. This is done only if the new networking solution is fully monitored, secured and can provide rapid revenues (short Return of Investment). For example, the NDN (Named Data Networking) solution is admitted as promising but still uncertain, thus making network operators reluctant to deploy it. Having a flexible environment would allow network operators to initiate the deployment of new network solutions at low cost and low risk. The virtualization techniques, appeared a few years ago, can help to provide such a flexible networking architecture. However, with it, emerge monitoring and security issues which should be solved. In this paper, we present our secure virtualized networking environment to deploy new functions and protocol stacks in the network, with a specific focus on the NDN use-case as one of the potential Future Internet technology. As strong requirements for a network operator, we then focus on monitoring and security components, highlighting where and how they can be deployed and used. Finally, we introduce our preliminary evaluation, with a focus on security, before presenting the test bed, involving end-users consuming real contents, that we will set up for the assessment of our approach.