Abstract Model-agnostic feature attribution techniques are used to explain the decisions of complex machine learning (ML) models including ensemble models, and deep neural networks (DNNs). However, since complex ML models perform best when trained on low-level features, the explanations generated by these algorithms are often not interpretable or usable by humans. Recently proposed model-agnostic methods that support the generation of human-interpretable explanations are impractical because they require a fully invertible transformation function that maps the model’s input features to human-interpretable features. While some practical human-interpretable explainability methods exist (e.g., concept-based methods), they typically require direct access to the model and are not fully model-agnostic. In this paper, we introduce Latent SHAP, a model-agnostic black-box feature attribution framework that provides human-interpretable explanations without necessitating a fully invertible transformation function. We validate the fidelity of Latent SHAP ’s explanations through quantitative faithfulness assessments on two controlled datasets—a self-generated artificial dataset and the dSprites dataset. Furthermore, we showcase the practical utility of Latent SHAP in various real-world scenarios across domains such as computer vision, natural language processing, and cybersecurity. Each domain involves complex models (ensembles, DNNs, and LLMs), where invertible transformation functions are not available.

We present a highly-scalable secure computation of graph algorithms, which hides all information about the topology of the graph or other input values associated with nodes or edges. The setting is where all nodes and edges of the graph are secret-shared between multiple servers, and a secure computation protocol is run between these servers. While the method is general, we demonstrate it in a 3-server setting with an honest majority, with either semi-honest security or full security. A major technical contribution of our work is replacing the usage of secure sort protocols with secure shuffles, which are much more efficient. Full security against malicious behavior is achieved by adding an efficient verification for the shuffle operation, and computing circuits using fully secure protocols. We demonstrate the applicability of this technology by implementing two major algorithms: computing breadth-first search (BFS), which is also useful for contact tracing on private contact graphs, and computing maximal independent set (MIS). We implement both algorithms, with both semi-honest and full security, and run them within seconds on graphs of millions of elements.

We assess the vulnerabilities of deep face recognition systems for images that falsify/spoof multiple identities simultaneously. We demonstrate that, by manipulating the deep feature representation extracted from a face image via imperceptibly small perturbations added at the pixel level using our proposed method, one can fool a face verification system into recognizing that the face image belongs to multiple different identities with a high success rate. One characteristic of the UAXs crafted with our method is that they are universal (identity-agnostic); they are successful even against identities not known in advance. For a certain deep neural network, we show that we are able to spoof almost all tested identities (99%), including those not known beforehand (not included in training). Our results indicate that a multiple-identity attack is a real threat and should be taken into account when deploying face recognition systems.

Protocols for secure multiparty computation (MPC) enable a set of mutually distrusting parties to compute an arbitrary function of their inputs while preserving basic security properties like privacy and correctness. The study of MPC was initiated in the 1980s where it was shown that any function can be securely computed, thus demonstrating the power of this notion. However, these proofs of feasibility were theoretical in nature and it is only recently that MPC protocols started to become efficient enough for use in practice. Today, we have protocols that can carry out large and complex computations in very reasonable time (and can even be very fast, depending on the computation and the setting). Despite this amazing progress, there is still a major obstacle to the adoption and use of MPC due to the huge expertise needed to design a specific MPC execution. In particular, the function to be computed needs to be represented as an appropriate Boolean or arithmetic circuit, and this requires very specific expertise. In order to overcome this, there has been considerable work on compilation of code to (typically) Boolean circuits. One work in this direction takes a different approach, and this is the SPDZ compiler (not to be confused with the SPDZ protocol) that takes high-level Python code and provides an MPC run-time environment for securely executing that code. The SPDZ compiler can deal with arithmetic and non-arithmetic operations and is extremely powerful. However, until now, the SPDZ compiler could only be used for the specific SPDZ family of protocols, making its general applicability and usefulness very limited. In this paper, we extend the SPDZ compiler so that it can work with general underlying protocols. Our SPDZ extensions were made in mind to enable the use of SPDZ for arbitrary protocols and to make it easy for others to integrate existing and new protocols. We integrated three different types of protocols, an honest-majority protocol for computing arithmetic circuits over a field (for any number of parties), a three-party honest majority protocol for computing arithmetic circuits over the ring of integers Z2n, and the multiparty BMR protocol for computing Boolean circuits. We show that a single high-level SPDZ-Python program can be executed using all of these underlying protocols (as well as the original SPDZ protocol), thereby making SPDZ a true general run-time MPC environment.In order to be able to handle both arithmetic and non-arithmetic operations, the SPDZ compiler relies on conversions from field elements to bits and back. However, these conversions do not apply to ring elements (in particular, they require element division), and we therefore introduce new bit decomposition and recomposition protocols for the ring over integers with replicated secret sharing. These conversions are of independent interest and utilize the structure of Z2n (which is much more amenable to bit decomposition than prime-order fields), and are thus much more efficient than all previous methods. We demonstrate our compiler extensions by running a complex SQL query and a decision tree evaluation over all protocols.

A variety of secure multiparty computation (MPC) protocols have been proposed up to now. Since their performance characteristics are incomparable, the most suitable MPC protocol may be completely different depending on the given computational task and environment. It is tedious work to compare all the possibility to choose the most suitable MPC. The paper " Generalizing the SPDZ Compiler For Other Protocols'' in this ACM-CCS 2018 shows a framework for adding MPC protocols to a development tool of MPC program called "SPDZ'', which enables to compare multiple protocols easily. This poster and demo show how this framework is useful for choosing the suitable protocol for given target computation and environment.

Secure multiparty computation enables a set of parties to securely carry out a joint computation of their private inputs without revealing anything but the output. In the past few years, the efficiency of secure computation protocols has increased in leaps and bounds. However, when considering the case of security in the presence of malicious adversaries (who may arbitrarily deviate from the protocol specification), we are still very far from achieving high efficiency. In this paper, we consider the specific case of three parties and an honest majority. We provide general techniques for improving efficiency of cut-and-choose protocols on multiplication triples and utilize them to significantly improve the recently published protocol of Furukawa et al. (ePrint 2016/944). We reduce the bandwidth of their protocol down from 10 bits per AND gate to 7 bits per AND gate, and show how to improve some computationally expensive parts of their protocol. Most notably, we design cache-efficient shuffling techniques for implementing cut-and-choose without randomly permuting large arrays (which is very slow due to continual cache misses). We provide a combinatorial analysis of our techniques, bounding the cheating probability of the adversary. Our implementation achieves a rate of approximately 1.15 billion AND gates per second on a cluster of three 20-core machines with a 10Gbps network. Thus, we can securely compute 212,000 AES encryptions per second (which is hundreds of times faster than previous work for this setting). Our results demonstrate that high-throughput secure computation for malicious adversaries is possible.

Secure multi-party computation (SMPC) is a cryptographic tool that enables a set of parties to jointly compute any function of their inputs while keeping the privacy of inputs. The paper "High Throughput Semi-Honest Secure Three-Party Computation with an Honest Majority" in this ACM CCS 2016 [4] presents a new protocol which its implementation carried out over 1,300,000 AESs per second and was able to support 35,000 login queries of Kerberos authentication per second. This poster/demo presents the design of the implementation and demonstrates the Kerberos authentication over here. The design will show how this high-throughput three-party computation can be done using simple servers. The demonstration proves that secure multiparty computation of Kerberos authentications in large organizations is now practical.

In this paper, we describe a new information-theoretic protocol (and a computationally-secure variant) for secure three-party computation with an honest majority. The protocol has very minimal computation and communication; for Boolean circuits, each party sends only a single bit for every AND gate (and nothing is sent for XOR gates). Our protocol is (simulation-based) secure in the presence of semi-honest adversaries, and achieves privacy in the client/server model in the presence of malicious adversaries. On a cluster of three 20-core servers with a 10Gbps connection, the implementation of our protocol carries out over 1.3 million AES computations per second, which involves processing over 7 billion gates per second. In addition, we developed a Kerberos extension that replaces the ticket-granting-ticket encryption on the Key Distribution Center (KDC) in MIT-Kerberos with our protocol, using keys/ passwords that are shared between the servers. This enables the use of Kerberos while protecting passwords. Our implementation is able to support a login storm of over 35,000 logins per second, which suffices even for very large organizations. Our work demonstrates that high-throughput secure computation is possible on standard hardware.