Comprehensive Notes on Machine Learning Applications, Malware Detection Techniques, and Large Language Model Analysis

Analyzing Hardware Based Malware Detectors

  • Motivators for choosing hardware-based malware detectors over software-based ones:
    • Software-based detection executes too slowly to classify malware in time.
  • Resilience of hardware-based malware detectors:
    • Operate at a lower system level, below the operating system.
    • Less susceptible to tampering by malware that targets software layers.

A Low Cost EDA-based Stress Detection Using Machine Learning

  • Main motivator for preferring EDA over questionnaires:
    • Questionnaires are subjective and time-consuming.
    • EDA determines stress using skin conductivity with high accuracy, saving time and resources.
  • Key advantage of using the RandomForest algorithm:
    • Robust to noisy data due to the use of multiple decision trees.

Fake News Detection using Machine Learning Algorithms

  • Main Reason to use machine learning in fake news detection:
    • Machine learning can automatically recognize patterns in large amounts of text.
  • Requirement for Machine Learning Models before they can classify fake news:
    • A large dataset of labeled real and fake articles

Efficient Utilization of Adversarial Training towards Robust Machine Learners and its Analysis

  • Most generalizable retrained classifier to other adversarial attacks (options given):
    • Not specified in the content. Mentioned attacks are Jacobian Saliency Map Attack, Fast Gradient Sign Method, Carlini and Wagner, and Deep Fool Attack.
  • Drawbacks of increasing the "complexity" of the adversarial attack method:
    • Increasing complexity tends to achieve its goal better (i.e. MIM~=BIM>FGSM).
    • Harder to implement (either takes more time or is more complex to implement).
    • Certain algorithms are robust against types of adversarial training (i.e. CW).

Pyramid: Machine Learning Framework to Estimate the Optimal Timing and Resource Usage of a High-Level Synthesis Design

  • What is not solved through Pyramid when measuring resource usage for FPGAs?
    • Options include: Longer processing times for larger designs, Inaccuracies across differing FPGA devices, Inefficient resource usage, Optimization when determining maximum throughput.
  • Main purpose of the Pyramid in FPGA design:
    • To accurately predict hardware performance and resource usage using machine learning.

Predictive Models for Hospital Readmission Risk: A Systematic Review of Methods

  • Significance of Area Under ROC Curve:
    • AUC = 0 is chance performance.
    • AUC = 0.5 is uninformed classifier.
    • AUC = 1 is perfect performance.
  • Improving results of a linear model with significant class imbalance:
    • Use a sample that has a similar distribution to the population.

DNN Model Architecture Fingerprinting Attack on CPU-GPU Edge Devices

  • How to use a fingerprinting attack to identify an AI model without hacking it:
    • By tracking RAM, CPU, GPU usage.
  • Difference between transferability and portability:
    • Transferability measures generalization to new model variants.
    • Portability measures how well the attack works across multiple platforms.

Survey of Machine Learning for Electronic Design Automation

  • Why are the majority of ML models for electronic design automation RL or NN models?
    • They are effective at handling high-dimensional combinatorial data in the design spaces, and can learn complex optimization using feedback systems.
  • Why could the non-linearity of ML models be an issue for designers and engineers using EDA tools?
    • The lack of interpretability can make debugging very difficult.

Pain Level Modeling of Intensive Care Unit Patients with Machine Learning Methods: An Effective Congeneric Clustering-based Approach

  • Major limitations on using machine learning for measuring pain:
    • There isn't a large specialized dataset to train off of.
    • Pain is still specific for individuals and groups, so a general model can only go so far.
  • Why would using machine learning to estimate pain be more desirable than just asking the patient?
    • Scale of 1-10 is vague
    • Patients may want to over/under report their pain
    • Some patients may have limited communication methods

R2AD: Randomization and Reconstructor-based Adversarial Defense for Deep Neural Networks

  • Two main components of the R2AD defense and how they work:
    • R2AD combines (1) Random Nullification (RNF), which randomly masks parts of the input to disrupt adversarial patterns, and (2) a Reconstructor (autoencoder), which attempts to recover the original input and computes reconstruction error to detect suspicious inputs.
    • Together, they reject adversarial examples and preserve clean inputs.
  • Trade-off of using the Random Nullification (RNF) layer:
    • RNF weakens adversarial perturbations by nullifying input features, but it may also remove important features from clean inputs, reducing model accuracy.
    • The defense is effective but requires careful parameter tuning to balance security and performance.

Adversarial Attack on Microarchitectural Events based Malware Detectors

  • Why are HMD systems fast and lightweight?
    • They use built-in CPU counters to monitor hardware events without analyzing code, so there's no extra overhead. Great for embedded devices.
  • How can an attacker use a separate thread to generate adversarial hardware performance events without modifying the original program?
    • By running a separate wrapper program that creates fake hardware activity like cache and branch misses.

Ensemble Learning for Effective Run-Time Hardware-Based Malware Detection: A Comprehensive Analysis and Classification

  • Primary goal of ensemble learning:
    • To combine multiple models to improve overall performance
  • Difference between bagging and boosting:
    • Bagging reduces variance, while boosting reduces bias.

Machine Learning Techniques in Wireless Sensor Network Based Precision Agriculture

  • Common ML techniques used for crop yield prediction:
    • Ensemble Learning
    • Bayesian Models
    • Support Vector Machines
    • Artificial Neural Networks
  • How is artificial intelligence (AI) most commonly used in modern agriculture?
    • Analyzing data to optimize irrigation, detect pests, and predict crop yields

TinyML Meets IoT: A Comprehensive Survey

  • Can TinyML do both training and inference?
    • No, typically only inference is performed on the device. Training is done offline on a server or cloud, and the compressed model is deployed to the device.
  • Primary objective of quantization in the optimization of TinyML models:
    • To make the model smaller and faster by using fewer bits for numbers.

An Overview of Privacy in Machine Learning

  • Primary difference between centralized and federated learning in terms of data storage:
    • Centralized learning gathers all data in one place; federated learning keeps data on local devices.
  • What is a membership inference attack?
    • Given a model, the adversary infers whether a data point was used in the training set.

A Machine Learning Model for Emotion Recognition from Physiological Signals

  • Which machine learning model achieved the highest accuracy in recognizing amusement, sadness, and neutral emotions using the selected physiological features?
    • Correct Answer: B. Support Vector Machine with Linear Kernel (SVML)
  • What is the main advantage of using physiological signals, such as GSR and PPG, for emotion recognition compared to facial expressions or speech?
    • Correct Answer: C. They are involuntary and harder to fake

Large Language Models for Code Analysis: Do {LLMs} Really Do Their Job?

  • What does the prompt construction strategy in this study reveal about how LLMs are best used in software engineering tasks?
    • Prompt phrasing and structure play a critical role in how LLMs interpret and perform analysis tasks
  • What is a key reason the authors evaluated LLMs on obfuscated code?
    • To examine whether LLMs are robust to real-world transformations that break code readability

Does Synthetic Data Generation of LLMs Help Clinical Text Mining?

  • Which of the following is the primary advantage of using synthetic data generated by a large language model (LLM) like ChatGPT to train a local model for clinical NLP tasks?
    • Answer: C. It reduces privacy risks by avoiding the upload of real patient information while providing labeled examples.
  • In a clinical relation extraction (RE) task, a model is fine-tuned on synthetic examples produced by an LLM. Which is the most realistic outcome?
    • Answer: C. The model performs well on a held-out test set of manually labeled examples

Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference

  • What are some limitations of traditional LLM benchmarks that Chatbot Arena addresses?
    • Traditional benchmarks often rely on static datasets and predefined answers, which may not capture the nuances of open-ended tasks or evolving real-world applications.
  • What is the significance of using pairwise comparison in Chatbot Arena?
    • Pairwise comparisons simplify the evaluation process by focusing on relative preferences between two models, making it easier to gather and interpret human judgments.