DLP API 可檢查文字,並根據各種預先定義和自訂的 infoType 偵測工具,識別及分類機密資訊。系統識別出個人資訊後,即可套用去識別化技術,例如遮蓋、遮罩或權杖化。您也可以使用 DLP API 封鎖關鍵字。輸入內容保護:將使用者提示或資料傳送至 Gemini 前,您可以先透過 DLP API 遮蓋或遮蔽任何私密資訊。這樣可避免模型處理或記錄敏感資料。輸出內容保護:如果 Gemini 可能會無意間生成或揭露私密資訊 (例如摘要含有 PII 的來源文件),DLP API 可以在輸出內容傳送給使用者前掃描內容。
AI 領域和濫用方法不斷演進,因此持續評估 AI 系統的安全性至關重要。定期評估有助於找出安全漏洞、評估緩解措施的成效、因應不斷變化的風險、確保符合政策和價值觀、建立信任,以及維持法規遵循狀態。為此,我們採用各種評估類型,包括開發評估、保證評估、紅隊測試、外部評估和基準測試。評估範圍應涵蓋內容安全性、品牌安全、相關性、偏誤和公平性、真實性,以及抵禦對抗性攻擊的穩定性。Vertex AI 的生成式 AI 評估服務等工具可協助您完成這些工作,並強調根據評估結果進行反覆改良,是負責任地開發 AI 的必要做法。
[[["容易理解","easyToUnderstand","thumb-up"],["確實解決了我的問題","solvedMyProblem","thumb-up"],["其他","otherUp","thumb-up"]],[["難以理解","hardToUnderstand","thumb-down"],["資訊或程式碼範例有誤","incorrectInformationOrSampleCode","thumb-down"],["缺少我需要的資訊/範例","missingTheInformationSamplesINeed","thumb-down"],["翻譯問題","translationIssue","thumb-down"],["其他","otherDown","thumb-down"]],["上次更新時間:2025-09-04 (世界標準時間)。"],[],[],null,["# Safety in Vertex AI\n\nGenerative AI models like Gemini require robust safety measures to\nmitigate risks such as generating harmful content, leaking sensitive\ninformation, or being misused. Google Cloud's Vertex AI platform\nprovides a suite of tools and practices to implement holistic safety for your\nGemini models.\n\nPotential safety risks and mitigation strategies\n------------------------------------------------\n\nWhen deploying Gemini models, it's crucial to identify and mitigate\nvarious potential risks. A proactive approach to understanding these risks\nallows for more effective implementation of safety measures. A multi-layered\napproach to safety is critical, as it can mitigate or prevent:\n\n- **Content risks:** These can include content that's harmful, profanity and sexualization, and violence and gore.\n- **Brand safety risks:** Generated content may not align with your brand's tone or values, it may endorse competitors or inappropriate products, or generate content that can result in reputational damage.\n- **Alignment risks:** Generated content may be irrelevant or inaccurate.\n- **Security and privacy risks:** Generated content may leak sensitive training data or prompts, or adversarial users may attempt to force the model to override safety protocols or behave in unintended ways.\n\nOur deployed models offer various features to address these potential issues:\n\n- The default model and non-configurable filters provide a general safety net.\n- [System instructions](/vertex-ai/generative-ai/docs/multimodal/safety-system-instructions) provide direct guidance to the model on preferred behavior and topics to avoid.\n- [Content filters](/vertex-ai/generative-ai/docs/multimodal/configure-safety-filters) allow you to set specific thresholds for common harm types.\n- [Gemini as a filter](/vertex-ai/generative-ai/docs/multimodal/gemini-for-filtering-and-moderation) offers an advanced, customizable checkpoint for complex or nuanced safety concerns that might be missed by the preceding layers or require more context-aware evaluation.\n- [DLP](/sensitive-data-protection/docs/sensitive-data-protection-overview#api) specifically addresses the critical risk of sensitive data leakage, in case the model has access to sensitive data. It also enables the ability to create custom block lists.\n\n### Available safety tools in Vertex AI for Gemini\n\nVertex AI offers several tools to manage the safety of your\nGemini models. Understanding how each works, their considerations, and\nideal use cases will help you build a tailored safety solution.\n\n### Continuous safety evaluation\n\nContinuous safety evaluation is crucial for AI systems, as the AI landscape and\nmisuse methods are constantly evolving. Regular evaluations help identify\nvulnerabilities, assess mitigation effectiveness, adapt to evolving risks,\nensure alignment with policies and values, build trust, and maintain compliance.\nVarious evaluation types, including development evaluations, assurance\nevaluations, red teaming, external evaluations, and benchmark testing, help\nachieve this. The scope of evaluation should cover content safety, brand safety,\nrelevance, bias and fairness, truthfulness, and robustness to adversarial\nattacks. Tools like Vertex AI's [Gen AI evaluation\nservice](/vertex-ai/generative-ai/docs/models/evaluation-overview) can assist in\nthese efforts, emphasizing that iterative improvements based on evaluation\nfindings are essential for responsible AI development."]]