Helmholtz Center for Information Security
CISPA – Helmholtz-Zentrum für InformationssicherheitNot a member yet
3406 research outputs found
Sort by
Get a Model! Model Hijacking Attack Against Machine Learning Models
Get PDFMachine learning (ML) has established itself as a cornerstone for various critical applications ranging from autonomous driving to authentication systems. However, with this increasing adoption rate of machine learning models, multiple attacks have emerged. One class of such attacks is training time attack, whereby an adversary executes their attack before or during the machine learning model training. In this work, we propose a new training time attack against computer vision based machine learning models, namely model hijacking attack. The adversary aims to hijack a target model to execute a different task than its original one without the model owner noticing. Model hijacking can cause accountability and security risks since a hijacked model owner can be framed for having their model offering illegal or unethical services. Model hijacking attacks are launched in the same way as existing data poisoning attacks. However, one requirement of the model hijacking attack is to be stealthy, i.e., the data samples used to hijack the target model should look similar to the model’s original training dataset. To this end, we propose two different model hijacking attacks, namely Chameleon and Adverse Chameleon, based on a novel encoder-decoder style ML model, namely the Camouflager. Our evaluation shows that both of our model hijacking attacks achieve a high attack success rate, with a negligible drop in model utility
SSLGuard: A Watermarking Scheme for Self-supervised Learning Pre-trained Encoders
Get PDFSelf-supervised learning is an emerging machine learning (ML) paradigm. Compared to supervised learning that leverages high-quality labeled datasets to achieve good performance, self-supervised learning relies on unlabeled datasets to pre-train powerful encoders which can then be treated as feature extractors for various downstream tasks. The huge amount of data and computational resources consumption makes the encoders themselves become a valuable intellectual property of the model owner. Recent research has shown that the ML model's copyright is threatened by model stealing attacks, which aims to train a surrogate model to mimic the behavior of a given model. We empirically show that pre-trained encoders are highly vulnerable to model stealing attacks. However, most of the current efforts of copyright protection algorithms such as fingerprinting and watermarking concentrate on classifiers.
Meanwhile, the intrinsic challenges of pre-trained encoder's copyright protection remain largely unstudied. We fill the gap by proposing SSLGuard, the first watermarking algorithm for pre-trained encoders. Given a clean pre-trained encoder, SSLGuard embeds a watermark into it and outputs a watermarked version. The shadow training technique is also applied to preserve the watermark under potential model stealing attacks. Our extensive evaluation shows that SSLGuard is effective in watermark injection and verification, and is robust against model stealing and other watermark removal attacks such as pruning and finetuning
A Truly Robust Signal Temporal Logic: Monitoring Safety Properties of Interacting Cyber-Physical Systems under Uncertain Observation
Get PDFSignal Temporal Logic is a linear-time temporal logic designed for classifying the time-dependent signals originating from continuous-state or hybrid-state dynamical systems according to formal specifications. It has been conceived as a tool for systematizing the monitoring of cyber-physical systems, supporting the automatic translation of complex safety specifications into monitoring algorithms, faithfully representing their semantics. Almost all algorithms hitherto suggested do, however, assume perfect identity between the sensor readings, informing the monitor about the system state and the actual ground truth. Only recently have Visconti et al. addressed the issue of inexact measurements, taking up the simple model of interval-bounded per-sample error that is unrelated, in the sense of chosen afresh, across samples. We expand their analysis by decomposing the error into an unknown yet fixed offset and an independent per-sample error and show that in this setting, monitoring of temporal properties no longer coincides with collecting Boolean combinations of state predicates evaluated in each time instant over best-possible per-sample state estimates, but can be genuinely more informative in that it infers determinate truth values for monitoring conditions that interval-based evaluation remains inconclusive about. For the model-free as well as for the linear model-based case, we provide optimal evaluation algorithms based on affine arithmetic and SAT modulo theory, solving over linear arithmetic. The resulting algorithms provide conclusive monitoring verdicts in many cases where state estimations inherently remain inconclusive. In their model-based variants, they can simultaneously address the issues of uncertain sensing and partial observation
Information Flow Guided Synthesis
Get PDFCompositional synthesis relies on the discovery of assumptions, i.e., restrictions on the behavior of the remainder of the system that allow a component to realize its specification.
In order to avoid losing valid solutions, these assumptions should be necessary conditions for realizability. However, because there are typically many different behaviors that realize the same specification, necessary behavioral restrictions often do not exist.
In this paper, we introduce a new class of assumptions for compositional synthesis, which we call information flow assumptions. Such assumptions capture an essential aspect of distributed computing, because components often need to act upon information that is available only in other components. The presence of a certain flow of information is therefore often a necessary requirement, while
the actual behavior that establishes the information flow is unconstrained.
In contrast to behavioral assumptions, which are properties of individual computation traces, information flow assumptions are hyperproperties, i.e., properties of sets of traces. We present a method for the automatic derivation of information-flow assumptions from a temporal logic specification of the system. We then provide a technique for the automatic synthesis of component implementations based on information flow assumptions. This provides a new compositional approach to the synthesis of distributed systems. We report on encouraging first experiments with the approach, carried out with the BoSyHyper synthesis tool
A framework for constructing Single Secret Leader Election from MPC
No full textThe emergence of distributed digital currencies has raised the need for a reliable consensus mechanism. In proof-of-stake cryptocurrencies, the participants periodically choose a closed set of validators, who can vote and append transactions to the blockchain. Each validator can become a leader with the probability proportional to its stake. Keeping the leader private yet unique until it publishes a new block can significantly reduce the attack vector of an adversary and improve the throughput of the network. The problem of Single Secret Leader Election (SSLE) was first formally defined by Boneh et al. in 2020.
In this work, we propose a novel framework for constructing SSLE protocols, which relies on secure multi-party computation (MPC) and satisfies the desired security properties. Our framework does not use any shuffle or sort operations and has a computational cost for N parties as low as O(N) of basic MPC operations per party. We improve the state-of-the- art for SSLE protocols that do not assume a trusted setup. Moreover, our SSLE scheme efficiently handles weighted elections. That is, for a total weight S of N parties, the associated costs are only increased by a factor of logS. When the MPC layer is instantiated with techniques based on Shamir’s secret-sharing, our SSLE has a communication cost of O(N2) which is spread over O(log N) rounds, and can tolerate up to t < N/2 of faulty nodes without restarting the protocol, and its security relies on DDH in the random oracle model. When the MPC layer is instantiated with more efficient techniques based on garbled circuits, our SSLE requires all parties to participate, up to N − 1 of which can be malicious, and its security is based on the random oracle model
SAPIC+: protocol verifiers of the world, unite!
Get PDFSymbolic security protocol verifiers have reached a high degree of automation and maturity. Today, experts can model real-world protocols, but this often requires model-specific encodings and deep insight into the strengths and weaknesses of each of those tools. With SAPIC+, we introduce a protocol verification platform that lifts this burden and permits choosing the right tool for the job, at any development stage. We build on the existing compiler from SAPIC to TAMARIN, and extend it with automated translations from SAPIC+ to PROVERIF and DEEPSEC, as well as powerful, protocol-independent optimizations of the existing translation. We prove each part of these translations sound. A user can thus, with a single SAPIC+ file, verify reachability and equivalence properties on the specified protocol, either using PROVERIF, TAMARIN or DEEPSEC. Moreover, the soundness of the translation allows to directly assume results proven by another tool which allows to exploit the respective strengths of each tool. We demonstrate our approach by analyzing various existing models. This includes a large case study of the 5G authentication protocols, previously analyzed in TAMARIN. Encoding this model in SAPIC+ we demonstrate the effectiveness of our approach. Moreover, we study four new case studies: the LAKE-EDHOC and the Privacy-Pass protocols, both under standardization, the SSH protocol with the agent-forwarding feature, and the recent KEMTLS protocol, a post-quantum version of the main TLS key exchange
Comparing User Perceptions of Anti-Stalkerware Apps with the Technical Reality
Get PDFEvery year an increasing number of users face stalkerware on their phones. Many of them are victims of intimate partner surveillance (IPS) who are unsure how to identify or remove stalkerware from their phones. An intuitive approach would be to choose anti-stalkerware from the app store. However, a mismatch between user expectations and the technical capabilities can produce an illusion of security and risk compensation behavior (i.e., the Peltzmann effect).
We compare users' perceptions of anti-stalkerware with the technical reality. First, we applied thematic analysis to app reviews to analyze user perceptions. Then, we performed a cognitive walkthrough of two prominent anti-stalkerware apps available on the Google PlayStore and reverse-engineered them to understand their detection features.
Our results suggest that users base their trust on the look and feel of the app, the number and type of alerts, and the apps' affordances. We also found that app capabilities do not correspond to the users' perceptions and expectations, impacting their practical effectiveness. We discuss different stakeholders' options to remedy these challenges and better align user perceptions with the technical reality
Replication: Stories as Informal Lessons about Security
Get PDFAnecdotal stories about security threats told to non-experts by friends, peers, or the media have been shown to be im- portant in forming mental models and secure behaviors. In 2012, Rader et al. conducted a survey (n=301) of security stories with a student sample to determine factors that influence security perceptions and behavior. We replicated this survey with a more diverse sample (n=299), including different age groups and educational backgrounds. We were able to confirm many of the original findings, providing further evidence that certain characteristics of stories increase the likelihood of learning and retelling. Moreover, we contribute new insights into how people learn from stories, such as that younger and higher educated people are less likely to change their thinking or be emotionally influenced by stories. We (re)discovered all of the threat themes found by Rader et al., suggesting that these threats have not been eliminated in the last decade, and found new ones such as ransomware and data breaches. Our findings help to improve the design of security advise and education for non-experts
Understanding Utility and Privacy of Demographic Data in Education Technology by Causal Analysis and Adversarial-Censoring
No full textEducation technologies (EdTech) are becoming pervasive due to their cost-effectiveness, accessibility, and scalability. They also experienced accelerated market growth during the recent pandemic. EdTech collects massive amounts of students’ behavioral and (sensitive) demographic data, often justified by the potential to help students by personalizing education. Researchers voiced concerns regarding privacy and data abuses (eg, targeted advertising) in the absence of clearly defined data collection and sharing policies. However, technical contributions to alleviating students’ privacy risks have been scarce. In this paper, we argue against collecting demographic data by showing that gender—a widely used demographic feature—does not causally affect students’ course performance: arguably the most popular target of predictive models. Then, we show that gender can be inferred from behavioral data; thus, simply leaving them out does not protect students’ privacy. Combining a feature selection mechanism with an adversarial censoring technique, we propose a novel approach to create a ‘private’ version of a dataset comprising of fewer features that predict the target without revealing the gender, and are interpretive. We conduct comprehensive experiments on a public dataset to demonstrate the robustness and generalizability of our mechanism
Input Invariants
Get PDFHow can we generate valid system inputs? Grammar-based fuzzers are highly efficient in producing syntactically valid system inputs. However, programs will often reject inputs that are semantically invalid. We introduce ISLa, a declarative specification language for context-sensitive properties of structured system inputs based on context-free grammars. With ISLa, it is possible to specify input constraints like “a variable has to be defined before it is used,” “the ‘file name’ block must be 100 bytes long,” or “the number of columns in all CSV rows must be identical.”
Such constraints go into the ISLa fuzzer, which leverages the power of solvers like Z3 to solve semantic constraints and, on top, handles quantifiers and predicates over grammar structure. We show that a few ISLa constraints suffice to produce 100% semantically valid inputs while still maintaining input diversity. ISLa can also parse and precisely validate inputs against semantic constraints.
ISLa constraints can be mined from existing input samples. For this, our ISLearn prototype uses a catalog of common patterns, instantiates these over input elements, and retains those candidates that hold for the inputs observed and whose instantiations are fully accepted by input-processing programs. The resulting constraints can then again be used for fuzzing and parsing