Researchers have developed a new data poisoning attack called BenchDoor that exploits incentives within machine learning benchmarks. This attack manipulates training or test data to implant backdoors while simultaneously improving benchmark scores. Experiments demonstrated that BenchDoor can achieve high success rates in implanting backdoors across various Q&A and text classification tasks, even after decontamination methods are applied. Separately, a foundation-style pretraining framework has been proposed for scalable and transferable distributed photovoltaic forecasting, which unifies historical data and weather predictions to outperform existing models and adapt to new sites. AI
IMPACT New attack vectors highlight security vulnerabilities in AI benchmark design, while advances in forecasting models could accelerate renewable energy integration.
RANK_REASON Two distinct research papers published on arXiv covering AI security and forecasting.
- alphaXiv
- arXiv
- BenchDoor
- CatalyzeX Code Finder for Papers
- DagsHub
- Gemma
- Gotit.pub
- Hugging Face
- Influence Flower
- photovoltaics
- Qwen
- ScienceCast
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
