We study a basic question about cryptographic watermarking for generative models: how reliable can a watermark remain when an adversary is allowed to corrupt the encoded signal? To address this question, we introduce a minimal coding abstraction that we call a zero-bit tamper-detection code. This is a secret-key procedure that samples a pseudorandom codeword and, given a candidate word, decides whether it should be treated as unmarked content or as the result of tampering with a valid codeword. It captures the two core requirements of robust watermarking: soundness and tamper detection. Within this abstraction we prove a sharp unconditional limit on robustness to independent symbol corruption. For an alphabet of size $q$, there is a critical corruption rate of $1-1/q$ such that no scheme with soundness, even relaxed to allow a fixed constant false positive probability on random content, can reliably detect tampering once an adversary can change more than this fraction of symbols. In particular, in the binary case no cryptographic watermark can remain robust if more than half of the encoded bits are modified. We also show that this threshold is tight by giving simple information-theoretic constructions that achieve soundness and tamper detection for all strictly smaller corruption rates. We then test experimentally whether this limit appears in practice by looking at the recent watermarking for images of Gunn, Zhao, and Song (ICLR 2025). We show that a simple crop and resize operation reliably flipped about half of the latent signs and consistently prevented belief-propagation decoding from recovering the codeword, erasing the watermark while leaving the image visually intact.

Despite the privacy-oriented design, federated learning (FL) remains vulnerable to privacy breaches due to the exposure of model update snapshots throughout training. Trusted Execution Environments (TEEs) offer hardware-based isolation to safeguard data and computations, providing a compelling foundation for privacy-preserving FL. However, the limited memory available in mainstream TEEs hinders the deployment of large-scale neural networks, such as GPT models, within these secure enclaves. To address this limitation, we propose Marshaled Learning, a novel FL framework that enables large neural network training across memory-constrained TEEs while ensuring strong privacy guarantees for both data and model owners. To achieve this, Marshaled Learning partitions a model into subnets and distributes them across clients according to their memory capacities, coordinating forward and backward passes across TEE-isolated environments. To mitigate the impact of heterogeneous data distributions and straggler clients, we introduce a dynamic knowledge propagation mechanism that facilitates cross-client learning and accelerates convergence. We present both theoretical convergence guarantees and empirical evaluations, demonstrating that Marshaled Learning outperforms existing FL methods by around 2% to 5% accuracy with much faster convergence rates. We also implement Marshaled Learning on commercial Azure Confidential VMs to prove its feasibility and show that it incurs only a 1 ~ 3× computational overhead compared to non-TEE settings, validating its practicality in real-world deployments.

Fine-tuning plays a crucial role in adapting models to downstream tasks with minimal training efforts. However, the rapidly increasing size of foundation models poses a daunting challenge for accommodating foundation model fine-tuning in most commercial devices, which often have limited memory bandwidth. Techniques like model sharding and tensor parallelism address this issue by distributing computation across multiple devices to meet memory requirements. Nevertheless, these methods do not fully leverage their foundation nature in facilitating the fine-tuning process, resulting in high computational costs and imbalanced workloads. We introduce a novel Distributed Dynamic Fine-Tuning (D2FT) framework that strategically orchestrates operations across attention modules based on our observation that not all attention modules are necessary for forward and backward propagation in fine-tuning foundation models. Through three innovative selection strategies, D2FT significantly reduces the computational workload required for fine-tuning foundation models. Furthermore, D2FT addresses workload imbalances in distributed computing environments by optimizing these selection strategies via multiple knapsack optimization. Our experimental results demonstrate that the proposed D2FT framework reduces the training computational costs by 40% and training communication costs by 50% with only 1% to 2% accuracy drops on the CIFAR-10, CIFAR-100, and Stanford Cars datasets. Moreover, the results show that D2FT can be effectively extended to recent LoRA, a state-of-the-art parameter-efficient fine-tuning technique. By reducing 40% computational cost or 50% communication cost, D2FT LoRA top-1 accuracy only drops 4% to 6% on Stanford Cars dataset. The extended version of this paper can be found in http://arxiv.org/abs/2504.12471.

Watermarking generative models consists of planting a statistical signal (watermark) in a model’s output so that it can be later verified that the output was generated by the given model. A strong watermarking scheme satisfies the property that a computationally bounded attacker cannot erase the watermark without causing significant quality degradation. In this paper, we study the (im)possibility of strong watermarking schemes. We prove that, under well-specified and natural assumptions, strong watermarking is impossible to achieve. This holds even in the private detection algorithm setting, where the watermark insertion and detection algorithms share a secret key, unknown to the attacker. To prove this result, we introduce a generic efficient watermark attack; the attacker is not required to know the private key of the scheme or even which scheme is used. Our attack is based on two assumptions: (1) The attacker has access to a "quality oracle" that can evaluate whether a candidate output is a high-quality response to a prompt, and (2) The attacker has access to a "perturbation oracle" which can modify an output with a nontrivial probability of maintaining quality, and which induces an efficiently mixing random walk on high-quality outputs. We argue that both assumptions can be satisfied in practice by an attacker with weaker computational capabilities than the watermarked model itself, to which the attacker has only black-box access. Furthermore, our assumptions will likely only be easier to satisfy over time as models grow in capabilities and modalities. We demonstrate the feasibility of our attack by instantiating it to attack three existing watermarking schemes for large language models: Kirchenbauer et al. (2023), Kuditipudi et al. (2023), and Zhao et al. (2023), and include preliminary results on vision-language models. The same attack successfully removes the watermarks planted by all schemes, with only minor quality degradation.

We introduce the concept of "universal" password model—a password model that, once pre-trained, can automatically adapt its guessing strategy based on the target system. To achieve this, the model does not need to access any plaintext passwords from the target credentials. Instead, it exploits users’ auxiliary information, such as email addresses, as a proxy signal to predict the underlying password distribution.Specifically, the model uses deep learning to capture the correlation between the auxiliary data of a group of users (e.g., users of a web application) and their passwords. It then exploits those patterns to create a tailored password model for the target system at inference time. No further training steps, targeted data collection, or prior knowledge of the community’s password distribution is required.Besides improving over current password strength estimation techniques, the model enables any end-user (e.g., system administrators) to autonomously generate tailored password models for their systems without the often unworkable requirements of collecting suitable training data and fitting the underlying machine learning model. Ultimately, our framework enables the democratization of well-calibrated password models to the community, addressing a major challenge in the deployment of password security solutions at scale.

Credential tweaking attacks use breached passwords to generate semantically similar passwords and gain access to victims’ services. These attacks sidestep the first generation of compromised credential checking (C3) services. The second generation of compromised credential checking services, called “Might I Get Pwned” (MIGP), is a privacy-preserving protocol that defends against credential tweaking attacks by allowing clients to query whether a password or a semantically similar variation is present in the server’s compromised credentials dataset. The desired privacy requirements include not revealing the user’s entered password to the server and ensuring that no compromised credentials are disclosed to the client.In this work, we formalize the cryptographic leakage of the MIGP protocol and perform a security analysis to assess its impact on the credentials held by the server. We focus on how this leakage aids breach extraction attacks, where an honest-but-curious client interacts with the server to extract information about the stored credentials. Furthermore, we discover additional leakage that arises from the implementation of Cloudflare’s deployment of MIGP. We evaluate how the discovered leakage affects the guessing capability of an attacker in relation to breach extraction attacks. Finally, we propose MIGP 2.0, a new iteration of the MIGP protocol designed to minimize data leakage and prevent the introduced attacks.

Secure aggregation is a cryptographic protocol that securely computes the aggregation of its inputs. It is pivotal in keeping model updates private in federated learning. Indeed, the use of secure aggregation prevents the server from learning the value and the source of the individual model updates provided by the users, hampering inference and data attribution attacks.

We investigate the security of split learning---a novel collaborative machine learning framework that enables peak performance by requiring minimal resource consumption. In the present paper, we expose vulnerabilities of the protocol and demonstrate its inherent insecurity by introducing general attack strategies targeting the reconstruction of clients' private training sets. More prominently, we show that a malicious server can actively hijack the learning process of the distributed model and bring it into an insecure state that enables inference attacks on clients' data. We implement different adaptations of the attack and test them on various datasets as well as within realistic threat scenarios. We demonstrate that our attack can overcome recently proposed defensive techniques aimed at enhancing the security of the split learning protocol. Finally, we also illustrate the protocol's insecurity against malicious clients by extending previously devised attacks for Federated Learning.

Learning useful representations from unstructured data is one of the core challenges, as well as a driving force, of modern data-driven approaches. Deep learning has demonstrated the broad advantages of learning and harnessing such representations.In this paper, we introduce a deep generative model representation learning approach for password guessing. We show that an abstract password representation naturally offers compelling and versatile properties that open new directions in the extensively studied, and yet presently active, password guessing field. These properties can establish novel password generation techniques that are neither feasible nor practical with the existing probabilistic and non-probabilistic approaches. Based on these properties, we introduce: (1) A general framework for conditional password guessing that can generate passwords with arbitrary biases; and (2) an Expectation Maximization-inspired framework that can dynamically adapt the estimated password distribution to match the distribution of the attacked password set.

Deep Learning has recently become hugely popular in machine learning for its ability to solve end-to-end learning systems, in which the features and the classifiers are learned simultaneously, providing significant improvements in classification accuracy in the presence of highly-structured and large databases.

We put forward a new framework that makes it possible to re-write or compress the content of any number of blocks in decentralized services exploiting the blockchain technology. As we argue, there are several reasons to prefer an editable blockchain, spanning from the necessity to remove inappropriate content and the possibility to support applications requiring re-writable storage, to "the right to be forgotten." Our approach generically leverages so-called chameleon hash functions (Krawczyk and Rabin, NDSS '00), which allow determining hash collisions efficiently, given a secret trapdoor information. We detail how to integrate a chameleon hash function in virtually any blockchain-based technology, for both cases where the power of redacting the blockchain content is in the hands of a single trusted entity and where such a capability is distributed among several distrustful parties (as is the case with Bitcoin). We also report on a proof-of-concept implementation of a redactable blockchain, building on top of Nakamoto's Bitcoin core. The prototype only requires minimal changes to the way current client software interprets the information stored in the blockchain and to the current blockchain, block, or transaction structures. Moreover, our experiments show that the overhead imposed by a redactable blockchain is small compared to the case of an immutable one.

Energy-harvesting techniques for low-power embedded devices are opening up new opportunities for the design and optimization of security protocols for Green Wireless Sensor Networks. In this paper, we focus on scenarios where the energy resources of nodes in the network are heterogeneous, and propose a network-level solution that leverages the heterogeneity of harvesting capabilities to reduce the energy consumption of performing costly security operations. Our proposed distributed protocol, called HELIOS (Harvesting- EnabLed computatIon Outsourcing Scheme), allows nodes with scarce energy availability to outsource resource-demanding cryptographic operations to nodes that are harvesting power in excess, resulting in a significant reduction of their energy consumption.

This work is motivated by a general question: can micro-scale energy-harvesting techniques be exploited to support low-cost standard security solutions on resource-constrained devices? We focus on guaranteeing integrity and authentication in Internet of Things (IoT) and Wireless Sensor Network (WSN) applications. In this article, we propose techniques to make ECDSA signatures low cost and implementable on resource-constrained devices. By combining precomputation techniques and energy-harvesting capabilities of modern sensor nodes, we achieve significant improvement over prior works. In addition, we show that the cost of ECDSA signatures can be reduced by up to a factor 10 by using harvesting-aware optimizations.

Remote data integrity checking (RDIC) enables a data storage server, say a cloud server, to prove to a verifier that it is actually storing a data owner's data honestly. To date, a number of RDIC protocols have been proposed in the literature, but most of the constructions suffer from the issue of a complex key management, that is, they rely on the expensive public key infrastructure (PKI), which might hinder the deployment of RDIC in practice. In this paper, we propose a new construction of identity-based (ID-based) RDIC protocol by making use of key-homomorphic cryptographic primitive to reduce the system complexity and the cost for establishing and managing the public key authentication framework in PKI-based RDIC schemes. We formalize ID-based RDIC and its security model, including security against a malicious cloud server and zero knowledge privacy against a third party verifier. The proposed ID-based RDIC protocol leaks no information of the stored data to the verifier during the RDIC process. The new construction is proven secure against the malicious server in the generic group model and achieves zero knowledge privacy against a verifier. Extensive security analysis and implementation results demonstrate that the proposed protocol is provably secure and practical in the real-world applications.

We provide a formal treatment of security of digital signatures against subversion attacks (SAs). Our model of subversion generalizes previous work in several directions, and is inspired by the proliferation of software attacks (e.g., malware and buffer overflow attacks), and by the recent revelations of Edward Snowden about intelligence agencies trying to surreptitiously sabotage cryptographic algorithms. The main security requirement we put forward demands that a signature scheme should remain unforgeable even in the presence of an attacker applying SAs (within a certain class of allowed attacks) in a fully-adaptive and continuous fashion. Previous notions---e.g., security against algorithm-substitution attacks introduced by Bellare et al. (CRYPTO '14) for symmetric encryption---were non-adaptive and non-continuous.

It is generally recognized that the network traffic generated by an individual acts as his biometric signature. Several tools exploit this fact to fingerprint and monitor users. Often, though, these tools access the entire traffic, including IP addresses and payloads. In general, this is not feasible on the grounds that both performance and privacy would be negatively affected. In reality, most ISPs convert user traffic into Net Flow records for a concise representation that does not include the payload. More importantly, a single IP address belonging to a large and distributed network is usually masked using Network Address Translation techniques, thus a few IP addresses may be associated to thousands of individuals (NAT'd IPs). We devised a new fingerprinting framework that overcomes these hurdles. Our system is able to analyze a huge amount of network traffic represented as Net Flows, with the intent to track people. It does so by accurately inferring when users are connected to the network and which IP addresses they are using, even though thousands of users are hidden behind NAT. Our prototype implementation was deployed and tested within an existing large metropolitan WiFi network serving about 200,000 users, with an average load of more than 1,000 users simultaneously connected behind 2 NAT'd IP addresses only. Our solution turned out to be very effective, with an accuracy greater than 90%. We also devised new tools and refined existing ones that may be applied to other contexts related to Net Flow analysis.

It is common nowadays for data owners to outsource their data to the cloud. Since the cloud cannot be fully trusted, the outsourced data should be encrypted. This however brings a range of problems, such as: How should a data owner grant search capabilities to the data users? How can the authorized data users search over a data owner's outsourced encrypted data? How can the data users be assured that the cloud faithfully executed the search operations on their behalf? Motivated by these questions, we propose a novel cryptographic solution, called verifiable attribute-based keyword search (VABKS). The solution allows a data user, whose credentials satisfy a data owner's access control policy, to (i) search over the data owner's outsourced encrypted data, (ii) outsource the tedious search operations to the cloud, and (iii) verify whether the cloud has faithfully executed the search operations. We formally define the security requirements of VA B K S and describe a construction that satisfies them. Performance evaluation shows that the proposed schemes are practical and deployable.

We introduce a model for provable data possession (PDP) that can be used for remote data checking: A client that has stored data at an untrusted server can verify that the server possesses the original data without retrieving it. The model generates probabilistic proofs of possession by sampling random sets of blocks from the server, which drastically reduces I/O costs. The client maintains a constant amount of metadata to verify the proof. The challenge/response protocol transmits a small, constant amount of data, which minimizes network communication. Thus, the PDP model for remote data checking is lightweight and supports large data sets in distributed storage systems. The model is also robust in that it incorporates mechanisms for mitigating arbitrary amounts of data corruption. We present two provably-secure PDP schemes that are more efficient than previous solutions. In particular, the overhead at the server is low (or even constant), as opposed to linear in the size of the data. We then propose a generic transformation that adds robustness to any remote data checking scheme based on spot checking. Experiments using our implementation verify the practicality of PDP and reveal that the performance of PDP is bounded by disk I/O and not by cryptographic computation. Finally, we conduct an in-depth experimental evaluation to study the tradeoffs in performance, security, and space overheads when adding robustness to a remote data checking scheme.

Many storage systems rely on replication to increase the availability and durability of data on untrusted storage systems. At present, such storage systems provide no strong evidence that multiple copies of the data are actually stored. Storage servers can collude to make it look like they are storing many copies of the data, whereas in reality they only store a single copy. We address this shortcoming through multiple-replica provable data possession (MR-PDP): A provably-secure scheme that allows a client that stores t replicas of a file in a storage system to verify through a challenge-response protocol that (1) each unique replica can be produced at the time of the challenge and that (2) the storage system uses t times the storage required to store a single replica. MR-PDP extends previous work on data possession proofs for a single copy of a file in a client/server storage system (Ateniese et al., 2007). Using MR-PDP to store t replicas is computationally much more efficient than using a single-replica PDP scheme to store t separate, unrelated files (e.g., by encrypting each file separately prior to storing it). Another advantage of MR-PDP is that it can generate further replicas on demand, at little expense, when some of the existing replicas fail.

Storage outsourcing is a rising trend which prompts a number of interesting security issues, many of which have been extensively investigated in the past. However, Provable Data Possession (PDP) is a topic that has only recently appeared in the research literature. The main issue is how to frequently, efficiently and securely verify that a storage server is faithfully storing its client's (potentially very large) outsourced data. The storage server is assumed to be untrusted in terms of both security and reliability. (In other words, it might maliciously or accidentally erase hosted data; it might also relegate it to slow or off-line storage.) The problem is exacerbated by the client being a small computing device with limited resources. Prior work has addressed this problem using either public key cryptography or requiring the client to outsource its data in encrypted form.

We introduce a model for provable data possession (PDP) that allows a client that has stored data at an untrusted server to verify that the server possesses the original data without retrieving it. The model generates probabilistic proofs of possession by sampling random sets of blocks from the server, which drastically reduces I/O costs. The client maintains a constant amount of metadata to verify the proof. The challenge/response protocol transmits a small, constant amount of data, which minimizes network communication. Thus, the PDP model for remote data checking supports large data sets in widely-distributed storage system.

A time-bound hierarchical key assignment scheme is a method to assign time-dependent encryption keys to a set of classes in a partially ordered hierarchy, in such a way that the key of a higher class can be used to derive the keys of all classes lower down in the hierarchy, according to temporal constraints.In this paper we design and analyze time-bound hierarchical key assignment schemes which are provably-secure and efficient. We first consider the unconditionally secure setting and we show a tight lower bound on the size of the private information distributed to each class. Then, we consider the computationally secure setting and obtain several results: We first prove that a recently proposed scheme is insecure against collusion attacks. Hence, motivated by the need for provably-secure schemes, we propose two different constructions for time-bound hierarchical key assignment schemes. The first one is based on symmetric encryption schemes, whereas, the second one makes use of bilinear maps. These appear to be the first constructions of time-bound hierarchical key assignment schemes which are simultaneously practical and provably-secure.

In 1998, Blaze, Bleumer, and Strauss (BBS) proposed an application called atomic proxy re-encryption , in which a semitrusted proxy converts a ciphertext for Alice into a ciphertext for Bob without seeing the underlying plaintext. We predict that fast and secure re-encryption will become increasingly popular as a method for managing encrypted file systems. Although efficiently computable, the wide-spread adoption of BBS re-encryption has been hindered by considerable security risks. Following recent work of Dodis and Ivan, we present new re-encryption schemes that realize a stronger notion of security and demonstrate the usefulness of proxy re-encryption as a method of adding access control to a secure file system. Performance measurements of our experimental file system demonstrate that proxy re-encryption can work effectively in practice.

We present a monitoring system for a dynamic network in which a set of domain nodes shares the responsibility for producing and storing monitoring information about a set of visitors. This information is stored persistently when the set of domain nodes grows and shrinks. Such a system can be used to store traffic or other logs for auditing or can be used as a subroutine for many applications to allow significant increases in functionality and reliability. The features of our system include authenticating visitors, monitoring their traffic through the domain, and storing this information in a persistent, efficient, and searchable manner. The storage process is O(log n)-competitive in the number of network messages with respect to an optimal offline algorithm; we show that this is as good as any online algorithm can achieve and significantly better than many commonly used strategies for distributed load balancing

We introduce a new cryptographic primitive, called insubvertible encryption, that produces ciphertexts which can be randomized without the need of any key material. Unlike plain universal re-encryption schemes, insubvertible encryption prevents against adversarial exploitation of hidden channels, by including certificates proving that the ciphertext can only be decrypted by authorized parties.The scheme can be applied to RFID tags, providing strong protection against tracing. This enables post-sale applications of manufacturer-issued RFID tags while preserving the privacy of consumers. The functionality required of the RFID tags is minimal, namely that they be re-writable (many-writable). No cryptographic capabilities are required of the tags themselves, as the readers perform all necessary computations.

In 1998, Blaze, Bleumer, and Strauss (BBS) proposed proxy re-signatures, in which a semi-trusted proxy acts as a translator between Alice and Bob. To translate, the proxy converts a signature from Alice into a signature from Bob on the same message. The proxy, however, does not learn any signing key and cannot sign arbitrary messages on behalf of either Alice or Bob. Since the BBS proposal, the proxy re-signature primitive has been largely ignored, but we show that it is a very useful tool for sharing web certificates, forming weak group signatures, and authenticating a network path.We begin our results by formalizing the definition of security for a proxy re-signature. We next substantiate the need for improved schemes by pointing out certain weaknesses of the original BBS proxy re-signature scheme which make it unfit for most practical applications. We then present two secure proxy re-signature schemes based on bilinear maps. Our first scheme relies on the Computational Diffie-Hellman (CDH) assumption; here the proxy can translate from Alice to Bob and vice-versa. Our second scheme relies on the CDH and 2-Discrete Logarithm (2-DL) assumptions and achieves a stronger security guarantee -- the proxy is only able to translate in one direction. Constructing such a scheme has been an open problem since proposed by BBS in 1998. Furthermore in this second scheme, even if the delegator and the proxy collude, they cannot sign on behalf of the delegatee. Both schemes are efficient and secure in the random oracle model.

This paper presents a new simple schemes for verifiable encryption of digital signatures. We make use of a trusted third party (TTP) but in an optimistic sense, that is, the TTP takes part in the protocol only if one user cheats or simply crashes. Our schemes can be used as primitives to build efficient fair exchange and certified e-mail protocols.

The Domain Name System (DNS) is a distributed database that allows convenient storing and retrieving of resource records. DNS has been extended to provide security services (DNSSEC) mainly through public-key cryptography. We propose a new approach to DNSSEC that may result in a signicantly more ecient protocol. We introduce a new strategy to build chains of trust from root servers to authoritative servers. The techniques we employ are based on symmetric-key cryptography. Keywords Domain Name System Security (DNSSEC), Authentication Protocols, Digital Signatures, Symmetric Encryption 1.

The increasing popularity and diversity of collaborative applications prompts a need for highly secure and reliable communication platforms for dynamic peer groups. Security mechanisms for such groups tend to be both expensive and complex and their integration with reliable group communication services presents a formidable challenge, This paper discusses some important integration issues, reports on our implementation experience and provides experimental results. Our approach utilizes distributed group key management developed by the Cliques project. We enhance it to handle processor and network faults (under a fail-stop or crash-and-recover model) and asynchronous membership events (such as joins, leaves, merges and network partitions). Our approach leverages the strong properties provided by the Spread group communication system, such as message ordering, clean failure semantics and a membership service. The result of this work is a secure group communications layer and an API that provide the application programmer with both standard group communication services and flexible security services.

Many modern computing environments involve dynamic peer groups. Distributed simulation, multiuser games, conferencing applications, and replicated servers are just a few examples. Given the openness of today's networks, communication among peers (group members) must be secure and, at the same time, efficient. This paper studies the problem of authenticated key agreement in dynamic peer groups with the emphasis on efficient and provably secure key authentication, key confirmation, and integrity. It begins by considering two-party authenticated key agreement and extends the results to group Diffie-Hellman (1976) key agreement. In the process, some new security properties (unique to groups) are encountered and discussed.

A fair exchange protocol allows two users to exchange items so that either each user gets the other's item or neither user does. In [2], verifiable encryption is introduced as a primitive that can be used to build extremely efficient fair exchange protocols where the items exchanged represent digital signatures. Such protocols may be used to digitally sign contracts.

Many modern computing environments involve dynamic peer groups. Distributed simulation, multi-user games, conferencing and replicated servers are just a few examples. Given the openness of today's networks, communication among group members must be secure and, at the same time, efficient. This paper studies the problem of authenticated key agreement in dynamic peer groups with the emphasis on efficient and provably secure key authentication, key confirmation and integrity. It begins by considering 2-party authenticated key agreement and extends the results to Group Diffie-Hellman key agreement. In the process, some new security properties (unique to groups) are discussed. 1 Introduction This paper is concerned with security services in the context of dynamic peer groups (DPGs). Such groups are common in many network protocol layers and in many areas of modern computing and the solution to their security needs, in particular key management, are still open research challenges [19]. Exa...