An official website of the United States government
Differential privacy is a strong notion for privacy that can be used to prove formal guarantees, in terms of a privacy budget, ϵ, about how much information is leaked by a mechanism. When used in privacy-preserving machine learning, the goal is typically to limit what can be inferred from the model about individual training records. However, the calibration of the privacy budget is not well understood. Implementations of privacy-preserving machine learning often select large values of ϵ in order to get acceptable utility of the model, with little understanding of the impact of such choices on meaningful privacy. Moreover, in scenarios where iterative learning procedures are used, relaxed definitions of differential privacy are often used which appear to reduce the needed privacy budget but present poorly understood trade-offs between privacy and utility. In this paper, we quantify the impact of these choices on privacy in experiments with logistic regression and neural network models. Our main finding is that there is no way to obtain privacy for free---relaxed definitions of differential privacy that reduce the amount of noise needed to improve utility also increase the measured privacy leakage. Current mechanisms for differentially private machine learning rarely offer acceptable utility-privacy trade-offs for complex learning tasks: settings that provide limited accuracy loss provide little effective privacy, and settings that provide strong privacy result in useless models.
more »
« less
This content will become publicly available on December 10, 2025
LAIP: Learned Adaptive Inspection Paths Using Offline Reinforcement Learning
Abstract—In many scenarios for informative path planning done by ground robots or drones, certain types of information are significantly more valuable than others. For example, in the precision agriculture context, detecting plant disease outbreaks can prevent costly crop losses. Quite often, there is a limit on the exploration budget, which does not allow for a detailed investigation of every location. In this paper, we propose Learned Adaptive Inspection Paths (LAIP), a methodology to learn policies that handle such scenarios by combining uniform sampling with close inspection of areas where high-value information is likely to be found. LAIP combines Q-learning in an offline reinforcement learning setting, careful engineering of the state representation and reward system, and a training regime inspired by the teacher-student curriculum learning model. We found that a policy learned with LAIP outperforms traditional approaches in low-budget scenarios.
more »
« less
- Award ID(s):
- 1931767
- PAR ID:
- 10579763
- Publisher / Repository:
- IEEE
- Date Published:
- ISSN:
- 0000
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
There have been great advances in bridge inspection damage detection involving the use of deep learning models. However, automated detection models currently fall short of giving an inspector an understanding of how the damage has progressed from one inspection to the next. The rate-of-change of the damage is a critical piece of information used by engineers to determine appropriate maintenance and rehabilitation actions to prevent structural failures. We propose a simple methodology for registering two bridge inspection videos or still images, collected at different stages of deterioration, so that trained model predictions may be directly measured and damage progression compared. The changes may be documented and presented to the inspector so that they may quickly evaluate key interest regions in the inspection video or image. Three approaches referred to as rigid, deformable, and hybrid image registration methods were experimentally tested and evaluated based on their ability to preserve the geometric characteristics of the referenced image. It was found in all experiments that the rigid, homography-based transformations performed the best for this application over a state-of-the-art deformable registration method, RANSAC-Flow.more » « less
-
Many infrastructure systems can be modeled as networks of components with binary states (intact, damaged). Information about components’ conditions is crucial for the maintenance process of the system. However, it is often impossible to collect information of all components due to budget constraints. Several metrics have been developed to assess the importance of the components in relation to maintenance actions: an important component is one that should receive high maintenance priority. Instead, in this paper we focus on the priority to be assigned for component inspections and information collection. We investigate metrics based on system level (global) and component level (local) decision making after inspection for networks with different topology, and compare these results with traditional ones. We then discuss the computational challenges of these metrics and provide possible approximation approaches.more » « less
-
Emerging technologies such as Augmented Reality (AR), have the potential to radically transform education by making challenging concepts visible and accessible to novices. In this project, we have designed a Hololens-based system in which collaborators are exposed to an unstructured learning activity in which they learned about the invisible physics involved in audio speakers. They learned topics ranging from spatial knowledge, such as shape of magnetic fields, to abstract conceptual knowledge, such as relationships between electricity and magnetism. We compared participants' learning, attitudes and collaboration with a tangible interface through multiple experimental conditions containing varying layers of AR information. We found that educational AR representations were beneficial for learning specific knowledge and increasing participants' self-efficacy (i.e., their ability to learn concepts in physics). However, we also found that participants in conditions that did not contain AR educational content, learned some concepts better than other groups and became more curious about physics. We discuss learning and collaboration differences, as well as benefits and detriments of implementing augmented reality for unstructured learning activities.more » « less
-
Once a programmer knows one language, they can leverage concepts and knowledge already learned, and easily pick up another programming language. But is that always the case? To understand if programmers have difficulty learning additional programming languages, we conducted an empirical study of Stack Overflow questions across 18 different programming languages. We hypothesized that previous knowledge could potentially interfere with learning a new programming language. From our inspection of 450 Stack Overflow questions, we found 276 instances of interference that occurred due to faulty assumptions originating from knowledge about a different language. To understand why these difficulties occurred, we conducted semi-structured interviews with 16 professional programmers. The interviews revealed that programmers make failed attempts to relate a new programming language with what they already know. Our findings inform design implications for technical authors, toolsmiths, and language designers, such as designing documentation and automated tools that reduce interference, anticipating uncommon language transitions during language design, and welcoming programmers not just into a language, but its entire ecosystem.more » « less
