Publications

Authorship verification is used to link texts written by the same author without needing a model per author, making it useful for deanonymizing users spreading text with malicious intent. Recent advances in Transformer-based language models hold great promise for author verification, though short context lengths and non-diverse training regimes present challenges for their practical application. In this work, we investigate the effect of these challenges in the application of a Cross-Encoder Transformer-based author verification system under multiple conditions. We perform experiments with four Transformer backbones using differently tuned variants of fanfiction data and found that our BigBird pipeline outperformed Longformer, RoBERTa, and ELECTRA and performed competitively against the official top ranked system from the PAN evaluation. We also examined the effect of authors and fandoms not seen in training on model performance. Through this, we found fandom has the greatest influence on true trials, pairs of text written by the same author, and that a balanced training dataset in terms of class and fandom performed the most consistently.

Private Automated Contact Tracing (PACT) was a collaborative team and effort formed during the beginning of the Coronavirus Disease 2019 (COVID-19) pandemic. PACT's mission was to enhance contact tracing in pandemic response by designing exposure-detection functions in personal digital communication devices that have maximal public health utility while preserving privacy. This report explains and discusses the use of automated exposure notification during the COVID-19 pandemic and to provide some recommendations for those who may try to design and deploy similar technologies in future pandemics.

In speaker verification, score calibration is employed to transform verification scores to log-likelihood ratios (LLRs) which are statistically interpretable. Conventional calibration techniques apply a global score transform. However, in condition-aware (CA) calibration, information conveying signal conditions is provided as input, allowing calibration to be adaptive. This paper explores a generative approach to condition-aware score calibration. It proposes a novel generative model for speaker verification trials, each which includes a trial score, a trial label, and the associated pair of speaker embeddings. Trials are assumed to be drawn from a discrete set of underlying signal conditions which are modeled as latent Categorical random variables, so that trial scores and speaker embeddings are drawn from condition-dependent distributions. An Expectation-Maximization (EM) Algorithm for parameter estimation of the proposed model is presented, which does not require condition labels and instead discovers relevant conditions in an unsupervised manner. The generative condition-aware (GCA) calibration transform is then derived as the log-likelihood ratio of a verification score given the observed pair of embeddings. Experimental results show the proposed approach to provide performance improvements on a variety of speaker verification tasks, outperforming static and condition-aware baseline calibration methods. GCA calibration is observed to improve the discriminative ability of the speaker verification system, as well as provide good calibration performance across a range of operating points. The benefits of the proposed method are observed for task-dependent models where signal conditions are known, for universal models which are robust across a range of conditions, and when facing unseen signal conditions.

Significant recent research has focused on applying deep neural network models to the problem of network traffic classification. At the same time, much has been written about the vulnerability of deep neural networks to adversarial inputs, both during training and inference. In this work, we consider launching backdoor poisoning attacks against an encrypted network traffic classifier. We consider attacks based on padding network packets, which has the benefit of preserving the functionality of the network traffic. In particular, we consider a handcrafted attack, as well as an optimized attack leveraging universal adversarial perturbations. We find that poisoning attacks can be extremely successful if the adversary has the ability to modify both the labels and the data (dirty label attacks) and somewhat successful, depending on the attack strength and the target class, if the adversary perturbs only the data (clean label attacks).

The COVID-19 pandemic placed unprecedented demands on the global public health systems for disease surveillance and contact tracing. Engineers and scientists recognized that it might be possible to augment the efforts of public health teams, if a system for automated digital contact tracing could be quickly devised and deployed to the population of smartphones. The Private Automated Contact Tracing (PACT) protocol was one of several digital contact tracing proposals offered worldwide. PACT’s mission—to preserve individuals’ privacy and anonymity while enabling them to quickly alert even nearby strangers of a likely risky exposure—was adopted by Google and Apple and realized in the Exposure Notifications (EN) service and API for mobile application development. The Exposure Notifications system, like many digital proximity tools, is based on Bluetooth signal strength estimation, and keeps much of the necessary information and computation on the smartphones themselves. It implemented a decentralized approach to contact tracing: the public health authority, and other governmental authorities, cannot access the records of an individual’s encounters with others; nor is physical location used or shared by the service. Although the service is available on most modern iOS and Android devices, it is not enabled by default; the individual must opt in to use a particular region’s implementation of the service, either by installing the regional app or by enrolling through a menu of regions in the operating system settings. Likewise, individuals must affirm their consent before the service can share anonymized infection status with the regional public health authority, and alert recent close contacts. The widespread availability of Exposure Notifications through Apple and Google’s platforms has made it a de facto world standard. Determining its accuracy and effectiveness as a public health tool has been a subject of intense interest. In July 2020, CDC’s Innovative Technologies Team designated MIT LL and the PACT team as trusted technical advisors on the deployment of private automated contact tracing systems as part of its overall public health response to COVID-19. The Innovative Technologies Team sought to answer the following key question regarding automated contact tracing: Does automated contact tracing have sufficient public health value that it is worthwhile to integrate it at scale into existing and evolving manual contact tracing systems? Rapidly rising caseloads necessitated parallel-path assessment activities of most mature systems at the time. When access to the Google and Apple Exposure Notifications system became available, MIT LL focused the assessment efforts on the systems being built and deployed. There were two immediate and significant challenges to observing and quantifying the performance of the system as a whole: first, the privacy preserving design decisions of PACT and the system implementers denied access to system-level performance metrics, and second, obtaining accurate “ground truth” data about risky encounters in the population, against which to measure the detector performance, would require an unacceptable level of effort and intrusion. Therefore, MIT LL designed a set of parallel research activities to decompose the problem into components that could be assessed quantifiably (Bluetooth sensor performance, algorithm performance, user preferences and behaviors), components that could be assessed qualitatively (potential cybersecurity risks, potential for malicious use), and components that could be modeled based on current and emergent knowledge (population-level effects). The MIT LL research team conducted early assessments of the privacy and security aspects of new EN app implementations and closely reviewed the available system code exercised by the apps, before conducting a series of phone-to-phone data collections both in the laboratory and in simulated real-world conditions. The data from these experiments fed into models and visualization tools created to predict and understand the risk score output of candidate “weights and thresholds” configurations for EN, i.e., to predict the performance of the system as-built against ground truth data for distance and duration of “exposure”. The data and performance predictions from this effort helped to inform the global and local community of practice in making configuration decisions, and can help to predict the performance of future versions of similar tools, or alternative implementations of the current system. We conducted a human factors and usability review of early app user interfaces and messaging from public health, and designed a follow-on large-scale survey to investigate questions about user trust and system adoption decisions. The results of the human factors, user trust, and adoption studies were used by U.S. public health jurisdictions to make adjustments to public-facing communications, and were shared with Apple and Google to improve the user interface. Information gathered from public health experts enabled us to better understand conventional contact tracing workflows and data streams, and we incorporated that information into an agent-based model of “hybrid” contact tracing plus Exposure Notifications. We then combined it with emerging reports on vaccination, mask effectiveness, social interaction, variant transmissibility, and our own data on the sensitivity and specificity of the Bluetooth “dose” estimator, to predict system-level effects under various conditions. Finally, we helped to establish a network of Exposure Notifications “practitioners” in public health, who surfaced desirable system-level key performance indicators (implemented during 2021 and 2022, in the Exposure Notifications Private Analytics system, or ENPA). At the conclusion of the program, many of the initial conditions of the pandemic had changed. The Exposure Notifications service was available to most of the world, but had only been deployed by 28 U.S. states and territories, and had not been adopted by much of the population in those regions. High case rates during the Omicron surge (December 2021 – January 2022) and newly available ENPA data offered the first hints at calculating “real” state-level performance metrics, but those data belong to the states and many are cautious about publishing. Although Google and Apple have stated that Exposure Notifications was designed for COVID-19, and will not be maintained in its current form after the pandemic ends, the public health and engineering communities show clear interest in using the “lessons learned” from Exposure Notifications and other similar solutions to preserve the capabilities developed and prepare better systems for future public health emergencies. The intent of this report is to document the work that has been completed, as well as to inform where the work could be updated or adapted to meet future needs.

Estimations of ankle moments can provide clinically helpful information on the function of lower extremities and further lead to insight on patient rehabilitation and assistive wearable exoskeleton design. Current methods for estimating ankle moments leave room for improvement, with most recent cutting-edge methods relying on machine learning models trained on wearable sEMG and IMU data. While machine learning eliminates many practical challenges that troubled more traditional human body models for this application, we aim to expand on prior work that showed the feasibility of using LSTM models by employing an ensemble of LSTM networks. We present an adaptive weighted LSTM ensemble network and demonstrate its performance during standing, walking, running, and sprinting. Our result show that the LSTM ensemble outperformed every single LSTM model component within the ensemble. Across every activity, the ensemble reduced median root mean squared error (RMSE) by 0.0017-0.0053 N. m/kg, which is 2.7 – 10.3% lower than the best performing single LSTM model. Hypothesis testing revealed that most reductions in RMSE were statistically significant between the ensemble and other single models across all activities and subjects. Future work may analyze different trajectory lengths and different combinations of LSTM submodels within the ensemble.

Phased array radars (PARs) are a promising observing technology, at the cusp of being available to the broader meteorological community. PARs offer near-instantaneous sampling of the atmosphere with flexible beam forming, multifunctionality, and low operational and maintenance costs and without mechanical inertia limitations. These PAR features are transformative compared to those offered by our current reflector-based meteorological radars. The integration of PARs into meteorological research has the potential to revolutionize the way we observe the atmosphere. The rate of adoption of PARs in research will depend on many factors, including (i) the need to continue educating the scientific community on the full technical capabilities and trade-offs of PARs through an engaging dialogue with the science and engineering communities and (ii) the need to communicate the breadth of scientific bottlenecks that PARs can overcome in atmospheric measurements and the new research avenues that are now possible using PARs in concert with other measurement systems. The former is the subject of a companion article that focuses on PAR technology while the latter is the objective here.

This dataset includes multimodal physiologic, flight performance, and user interaction data streams, collected as participants performed virtual flight tasks of varying difficulty. In virtual reality, individuals flew an "Instrument Landing System" (ILS) protocol, in which they had to land an aircraft mostly relying on the cockpit instrument readings. Participants were presented with four levels of difficulty, which were generated by varying wind speed, turbulence, and visibility. Each of the participants performed 12 runs, split into 3 blocks of four consecutive runs, one run at each difficulty, in a single experimental session. The sequence of difficulty levels was presented in a counterbalanced manner across blocks. Flight performance was quantified as a function of horizontal and vertical deviation from an ideal path towards the runway as well as deviation from the prescribed ideal speed of 115 knots. Multimodal physiological signals were aggregated and synchronized using Lab Streaming Layer. Descriptions of data quality are provided to assess each data stream. The starter code provides examples of loading and plotting the time synchronized data streams, extracting sample features from the eye tracking data, and building models to predict pilot performance from the physiology data streams.

During a pandemic such as COVID-19, non-pharmaceutical interventions (NPIs) can help protect public health; however, it is not always clear which actions will have the greatest positive impact, or what the trade-offs are between different options. Exposure Notification (EN) was introduced as a prevention measure during the COVID-19 pandemic to supplement traditional contact tracing activities. To predict the estimated impacts of EN, a model for "simulation of automated exposure notification" (SimAEN) was developed by researchers at MIT Lincoln Laboratory (MIT LL) with CDC funding [2]. The model was published through an accessible web interface, available for use by the general public at https://SimAEN.philab.cdc.gov/.

Pre-training on time series poses a unique challenge due to the potential mismatch between pre-training and target domains, such as shifts in temporal dynamics, fast-evolving trends, and long-range and short cyclic effects, which can lead to poor downstream performance. While domain adaptation methods can mitigate these shifts, most methods need examples directly from the target domain, making them suboptimal for pre-training. To address this challenge, methods need to accommodate target domains with different temporal dynamics and be capable of doing so without seeing any target examples during pre-training. Relative to other modalities, in time series, we expect that time-based and frequency-based representations of the same example are located close together in the time-frequency space. To this end, we posit that time-frequency consistency (TF-C) — embedding a time-based neighborhood of a particular example close to its frequency-based neighborhood and back—is desirable for pre-training. Motivated by TF-C, we define a decomposable pre-training model, where the self-supervised signal is provided by the distance between time and frequency components, each individually trained by contrastive estimation. We evaluate the new method on eight datasets, including electrodiagnostic testing, human activity recognition, mechanical fault detection, and physical status monitoring. Experiments against eight state-of-the-art methods show that TF-C outperforms baselines by 15.4% (F1 score) on average in one-to-one settings (e.g., fine-tuning an EEG-pretrained model on EMG data) and by up to 8.4% (F1 score) in challenging one-to-many settings (e.g., fine-tuning an EEG-pretrained model for either hand-gesture recognition or mechanical fault prediction), reflecting the breadth of scenarios that arise in real-world applications. The source code and datasets are available at https://anonymous.4open.science/r/TFC-pretraining-6B07.