大约两百年前,法国人尼埃普斯将一块涂有感光材料的白镴板放进暗箱,对准窗外。
光线穿过镜头,落在这块金属板上。经过漫长的曝光,窗外的风景留下了一些模糊的明暗。这张《窗外的勒格拉景色》,通常被视为现存最早的相机照片。
此后近两百年,摄影工艺经历了无数次变化。其中的核心并没有变——真实存在的东西反射或者发出光线,光穿过镜头,在感光介质上留下痕迹。
「照片」的词源,正是「光的书写/光的图」(photo-graph)。这个词本身就隐含着一层意思:画面里的那一刻,曾经真实发生过。

但也是在摄影术诞生后的两个世纪里,照片对于「真实」的映射实际上早已被消解。
如果你看过电影《寂静的朋友》,里面那种人工修饰底片的工艺,在 19 世纪末期就已流行开来。数码时代来了,Photoshop 的事情则更无需篇幅赘述;而到了生成式 AI 浪潮骤起的今天,一张「看起来真实」照片与现实之间那种理所当然的联系,早已荡然无存。
笔者怎么证明,一张照片真的来自现实世界?
作为世界上最多人使用的拍照设备的生产商,苹果给出了自己的答案——Apple Reference Image (苹果参考图像)。
把「水印」刻进传感器
给图片标记来源,并不是什么新鲜事。最直观的办法,是在 AI 生成的图片上留下可见标识。豆包的水印就是人们熟悉的例子。
可见水印也有一个显著的弱点,裁切、修图,甚至只需要再交给另一个 AI,就能让它从画面中消失。
近几年,谷歌推出的 SynthID 工艺,在 AI 防伪和 AI 鉴别领域流行了起来,涵盖谷歌自身以及 Anthropic、OpenAI 等公司都在使用。这是一整套包含文本、图像等各种模态的「隐水印」工艺,具体在图片上,也是将肉眼不可辨别的数字水印嵌入图片。
不过它仍然有识别上的局限性,例如一般人在看到一张照片、一段文字的时候,是不懂、也不会使用 SynthID 校验工具来验证真伪的。
既然已经越来越难判断一张图片是真是假,与其继续给 AI 生成的内容打上「假」的标记,另一种思路是:
给真正由相机拍下来的照片,留下「真」的证明。
这也正是 Apple Reference Image 的核心思路。
当然,给真实照片打上一枚「水印」,也不是在照片上盖一个「Shot on iPhone」的水印那么简单。
毕竟笔者凭什么相信,盖章之前的照片就是真的?假如一张照片已经完成处理、生成了 JPEG,平台才给它附上一份无法伪造的凭证。那这份凭证最多只能证明,照片在盖章之后没有被改动。至于盖章之前,它有没有被替换、有没有经过篡改,凭证本身并不能回答。
要让这份证明真正可信,必须把信任的起点继续向前推。一直推到数字摄影最接近现实世界的地方——
传感器。
当光线穿过镜头,抵达相机传感器,并在那里被转换成数字信号,现实世界便第一次以像素数据的形式进入手机。
苹果挑选在这里建立信任的起点。
笔者将它称为「水印」,但实际上是一份由传感器为捕获数据给出的密码学签名。iPhone 18 Pro 和 iPhone 18 Pro Max 启动参考模式后,主摄传感器会进入专门的安全捕获模式。
传感器捕获原始像素,随即用内部私钥对这些原始数据开展签名。硬件机制还会阻止传感器固件在签名前修改这些数据。
这样,签名保护的便不止是传感器输出了什么,还涵盖它输出的图片数据,与实际捕获结果之间的关系。
从一张照片,到一条信任链
苹果绝对不是可信拍摄工艺的发明者,此前传统相机厂商已经尝试为照片附上可验证的来源凭证。徕卡推出的 Leica Content Credentials(徕卡内容凭证)就是一个例子。
启动能力后,相机会利用机内的专用安全硬件,为生成的照片文件附上经过密码学签名的凭证。它使用 C2PA 标准,可以记录照片来源以及后续的修改。
值得关注的是这个凭证生成的时间节点:是在 JPEG 照片文件生成时建立的。
从传感器捕获原始像素,到文件生成,图像已经走过一系列处理步骤。
如果数据在签名前遭到替换或者篡改,最终文件即使取得有效签名,也未必忠实于传感器最初捕获的画面。
这也是苹果试图借助 Reference Image 进一步回答的难题:如何建立一条从原始捕获,到最终成片,都可以接受验证的信任链?
首先,平台得确认签名的传感器确实属于这台 iPhone。
在生产阶段,传感器会生成专属密钥。私钥留在传感器内部,公钥则用于验证签名。
iPhone 的安全隔区 (Secure Enclave) 也会建立独立的签名身份。苹果借助设备硬件清单,把两者记录为同一台 iPhone 的部件。
到了验证时,平台既能检查像素的传感器签名,也能核对这颗传感器与安全隔区是否属于同一台出厂设备——只有两者相互吻合,设备身份这一环才能借助验证。
当然,笔者也不能完全保证「从传感器进入」就是绝对安全的,苹果并没有假设传感器本身永远不会被攻破。
假设攻击者成功破解某颗传感器,获取了其中的私钥,就有可能利用这把密钥,为伪造的图像数据生成有效签名。
因而,Reference Image 还外观了一套撤销机制:一旦发现某颗传感器的安全性受到破坏,苹果就可以撤销与之关联的照片认证,让这些照片不再被视为具有可信来源。
换句话说,即使信任链的起点遭到攻破,苹果仍然保留了让相关照片失去有效认证的能力,不会让一把泄露的私钥永久成为伪造照片的通行证。
说真的,这么干的现实意义不大,但苹果就是把照片的设备级加密验真,做到了这么决。
设备身份确认之后,还得知道照片是什么时候拍的。
普通照片的 EXIF 会记录拍摄时间,但它通常依赖设备时钟。使用者可以手动修改时间,遭到入侵的操作平台也可能给出虚假的时间信息。
因而,Reference Image 还需要一份独立于平台时钟的时间依据。
iPhone 会定期从苹果的密码学时间戳服务获取经过签名的时间凭证,全球平均约每 15 分钟一次;拍摄完成后,还会再请求一份新的时间凭证。
借助两个可信的时间点,把拍摄行为夹在中间。苹果无法证明照片是在具体哪一秒拍下的,但可以证明,它产生于这两个经过验证的时间点之间。
换句话说,Reference Image 给出的是一个「相对模糊,但绝对可信」的时间戳。即使拍摄时没有网络,后一份时间凭证也可以在恢复连接后补齐。
至此,传感器捕获的原始像素、相关元数据、传感器与安全隔区的签名、设备硬件清单,以及可信时间信息,会被打包成一份安全数字底片 (Secure Digital Negative),以 DNG 格式保存在 iPhone 中。
但底片还不是笔者最终看到的照片。
RAW 到 JPEG,还要经过一系列图像处理。如果成片过程遭到篡改,前面建立的可信关系就可能在最后毁于一旦。
因而,这份底片还需要在可信的环境里完成「冲洗」。苹果把这一步交给了私有云计算 Private Cloud Compute(PCC)。
当我挑选生成 Reference Image 时,PCC 会先核对原始像素、签名、证书、设备身份和时间信息,确认它们可以组成一条完整且相互吻合的证据链。验证借助后,PCC 才会处理 RAW 数据,生成最终的 JPEG,下载到使用者设备本地。
这套流程真的太「复古」了:它就好比你真的拍了一卷胶片(对应苹果的安全数字底片),然后把胶片送到冲印店(对应苹果的 PCC 服务器)去完成冲洗。
不过笔者又凭什么相信,苹果的云端服务是「可信的环境」?
苹果的应对办法,是让云端服务本身也成为可以被验证的对象,不只是依靠苹果的单维度背书。
PCC 正式部署的软件版本,会被记录在受到密码学保护的透明日志中,对应的二进制文件也会向外部安全研究人员公开;上传底片之前,iPhone 还会确认目标服务器运行的是日志中记录的版本。
换句话说,苹果不仅要证明照片的来源可信,还要让负责验证照片的程序本身,也有据可查。
经过苹果签名服务的最后一道签名,笔者得到了一张可以独立验证来源的 Reference Image。至此,整条信任链终于闭合。
还有一个有意思的细节,PCC 的验证并不止于照片本身,它再次把目光放回这条信任链的源头,也就是传感器。
除了核对密码学签名,PCC 还会额外运行一个神经网络,检查图像是否具有苹果相机传感器原始输出应有的物理特征,并生成一份可信度评分。
这些评分会与照片的唯一标识符、传感器 ID 等信息一起交给苹果的撤销服务,用于持续评估传感器的可信状态。如果某颗传感器持续产生了异常的图像数据,苹果就可以据此发现潜在的安全风险,并在必要时撤销与这颗传感器关联的照片认证。
这样一来,整个验证过程也形成了一个循环:传感器借助密码学签名,为照片给出可信的来源证明;而照片本身,又成为苹果反过来检查传感器是否仍然可信的依据。
不过,当笔者试图为照片建立可信的来源时,还有一个容易被忽略的难题。
前面提到,每颗相机传感器都有自己的专属密钥,苹果也能借助设备硬件清单,确认照片来自哪台 iPhone。
那么,如果有人拿到多张 Reference Image,能否借助其中的传感器身份信息,判断它们是否来自同一台设备,进而追踪摄影者的活动?
对普通使用者来说,这或许只是一个隐私难题。不过,对于身处冲突地区的战地记者,或者正在调查敏感事件的摄影者来说,暴露拍摄设备的身份,可能意味着更加直接的人身安全风险。
因而,苹果没有把传感器的内部身份凭证公开在最终图像里。对于外部验证者来说,他们只能看到苹果为最终照片给出的签名,形象的来说只能看见苹果为照片盖上了一枚「经过可信相机捕获」的印章。
一种很「苹果」的机型逻辑
当然,Reference Image 也有它的边界。
举个最直接的例子:笔者完全可以用 iPhone 拍摄电脑显示屏上 AI 生成的假图。即使最终照片借助了全部验证,也只能证明相机确实捕获了显示屏上的画面——并不能证明画面中的事情曾经真实发生过。
换句话说,Reference Image 可以为照片的拍摄来源给出可信的证明,却无法保证照片所展现的内容一定真实。
苹果在官方文档中,也为这项工艺划定了一个明确克制的证明范围:
一颗可信的 iPhone 相机传感器,在某个经过验证的时间范围内捕获了这些光学数据,而这些数据随后经过了可验证的处理。
这也意味着,在 AI 时代,笔者还是无法简单地证明「画面里所讲的故事是真实的」;但至少,苹果的思路可以让一张照片多了一份关于自身来历的真实记录。
面对越来越难以辨别来源的 AI 图片,不同公司正在尝试不同的解法。
一家生成式 AI 公司,最自然的做法是在自己生成的内容上留下标记。Google DeepMind 的 SynthID ,在 AI 生成的图片中嵌入人眼难以察觉、但可以被专门工具识别的数字水印。
一家传统相机厂商,则更容易从拍摄设备本身出发。前面提到的徕卡 M11-P,就是借助机内安全硬件,为相机生成照片内容凭证。
这些思路的切入点不同,不过它们都从自己可以控制的环节开始,为图片建立可信的来源记录。
现在笔者再回过头来看 Apple Reference Image,就能发现这是一项格外「苹果式」的能力机型。
为了让一张照片具备可信的来源,苹果让传感器负责证明原始像素的捕获,让安全隔区和设备硬件清单确认设备身份,再借助可信时间戳确定拍摄发生的时间范围,最后交由 Private Cloud Compute 验证并生成最终的照片。
这套流程里,每一步都需要不同环节的配合。苹果可以把它们串起来,一个关键的前提是:从相机硬件、操作平台到云端服务,这些环节都在苹果自己的强大的掌控之中。
这让我想起苹果当年做灵动岛的方式。
显示屏上的挖孔原本只是一个硬件结构,苹果让平台动画、通知和应用状态围绕它重新组织,最后把它变成了可以交互的界面。
灵动岛的成立,依赖于硬件外观、平台交互和应用能力之间的协同。苹果把这些原本属于不同工艺层级的能力,放在同一套机型逻辑里考虑,最终形成了一项完整的能力。
Reference Image 与灵动岛虽然是两种完全不同的能力,但它们背后有着相似的机型思路。
在灵动岛上,目标是让显示屏上的硬件结构成为自然的交互界面;在 Reference Image 上,目标则是让一张照片从被捕获的那一刻起,就具备可以验证的来源记录。
它或许只是苹果应对 AI 时代的一次尝试。但从一个具体的难题出发,再用整个机型体系的能力去应对它,这样的应对思路,仔细想想好像确实 Only Apple Can Do。
如果你对 Apple Reference Image 背后的工艺细节感兴趣,也可以阅读苹果官方发布的参考文章:
https://security.apple.com/blog/apple-reference-image/
本文来源:https://www.ifanr.com/1681780
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
About two hundred years ago, the Frenchman Nicéphore Niépce placed a sheet of pewter coated with light-sensitive material inside a camera obscura, pointing it out the window.
Light passed through the lens and fell onto the metal plate. After an extremely long exposure, the scene outside the window left behind some faint patterns of light and shadow. This image, "View from the Window at Le Gras," is generally regarded as the earliest surviving photograph taken with a camera.
In the nearly two centuries since, photographic technology has undergone countless changes. Yet the core has never changed: real objects reflect or emit light, light passes through the lens, and leaves traces on a light-sensitive medium.
The etymology of "photograph" is precisely "writing with light / drawing with light" (photo-graph). The word itself carries the implication: that moment in the frame truly happened at some point.
But in the two centuries since the birth of photography, the mapping of photographs to "reality" has long since been dismantled.
If you have seen the movie "The Silent Friend," that kind of manually retouched negative process was already popular in the late 19th century. In the digital age, Photoshop needs no lengthy explanation; and today, with the sudden rise of generative AI, the once self-evident connection between a "realistic-looking" photo and reality has all but vanished.
How can we prove that a photo actually came from the real world?
Apple, the manufacturer of the world's most widely used camera devices, has given its own answer: Apple Reference Image.
Image | Apple
Engraving the "watermark" into the sensor
Marking the origin of images is nothing new. The most straightforward approach is to leave visible markers on AI-generated images. Doubao's watermark is a familiar example.
Visible watermarks also have a notable weakness: cropping, editing, or even just passing the image through another AI can make them disappear from the frame.
In recent years, Google's SynthID has become popular in the fields of AI anti-counterfeiting and AI authentication, being used by Google itself as well as companies such as Anthropic and OpenAI. It is a complete set of "invisible watermark" technologies covering text, images, and other modalities. For images specifically, it embeds imperceptible digital watermarks into the picture.
However, it still has limitations in verification. For example, ordinary people who come across a photo or a piece of text would neither understand nor use SynthID verification tools to confirm its authenticity.
Since it has become increasingly difficult to judge whether an image is real or fake, instead of continuing to mark AI-generated content with "fake" labels, another line of thinking is:
Leaving a "proof of authenticity" on photographs genuinely captured by cameras.
This is precisely the core idea behind Apple Reference Image.
Of course, stamping a "watermark" onto a real photo is not as simple as putting a "Shot on iPhone" mark on the image.
After all, why should we trust that a photo was real before the stamp was applied? Suppose a photo has already been processed and saved as a JPEG, and only then the platform attaches an unforgeable credential to it. At most, that credential can prove that the photo was not altered after the stamp was applied. As for what happened before the stamp — whether it was replaced or tampered with — the credential itself cannot answer that.
To make this proof truly trustworthy, the starting point of trust must be pushed further back. Back to the place where digital photography is closest to the real world:
The sensor.
When light passes through the lens, reaches the camera sensor, and is converted into digital signals there, the real world enters the phone for the first time in the form of pixel data.
Apple chooses to establish the starting point of trust here.
We call it a "watermark," but in reality it is a cryptographic signature given by the sensor to the captured data. Once the iPhone 18 Pro and iPhone 18 Pro Max have Reference Mode enabled, the main camera sensor enters a dedicated secure capture mode.
The sensor captures the raw pixels and immediately uses an internal private key to sign this raw data. Hardware mechanisms also prevent sensor firmware from modifying this data before signing.
In this way, the signature protects not only what the sensor outputs, but also the relationship between the image data it outputs and the actual captured result.
From a single photo to a chain of trust
Apple is certainly not the inventor of trustworthy capture technology. Traditional camera manufacturers have already attempted to attach verifiable provenance credentials to photos. Leica's Leica Content Credentials is one example.
Image | Leica
Once the capability is enabled, the camera uses its dedicated in-camera security hardware to attach a cryptographically signed credential to the generated photo file. It uses the C2PA standard and can record the photo's origin and subsequent modifications.
What is worth noting is the timing of when this credential is generated: it is established when the JPEG photo file is generated.
From the sensor capturing raw pixels to the generation of the file, the image has already gone through a series of processing steps.
If the data was replaced or tampered with before signing, the final file — even if it carries a valid signature — may not faithfully reflect what the sensor originally captured.
This is also the challenge Apple is trying to further address with Reference Image: how do you establish a chain of trust, from initial capture to final image, that can be verified at every step?
First, the platform must confirm that the signed sensor truly belongs to that particular iPhone.
During the production phase, the sensor generates a dedicated key pair. The private key remains inside the sensor, while the public key is used to verify the signature.
The iPhone's Secure Enclave also establishes an independent signing identity. Apple uses the device hardware inventory to record both as components of the same iPhone.
At verification time, the platform can check both the sensor signature of the pixels and whether that sensor and the Secure Enclave belong to the same factory-original device — only when the two match each other can the device identity link pass verification.
Of course, we cannot fully guarantee that "entering through the sensor" is absolutely safe. Apple does not assume that the sensor itself can never be compromised.
Suppose an attacker successfully cracks a sensor and extracts its private key; they could use that key to generate valid signatures for forged image data.
Therefore, Reference Image also includes a revocation mechanism: once it is discovered that a sensor's security has been compromised, Apple can revoke the photo authentication associated with it, so those photos are no longer regarded as having a trusted source.
In other words, even if the starting point of the chain of trust is broken, Apple retains the ability to strip valid authentication from related photos, ensuring that a leaked private key does not permanently become a pass for forged images.
To be honest, the practical significance of doing all this may be limited, but Apple has taken device-level cryptographic photo authentication to that level of determination.
After confirming device identity, it is also necessary to know when the photo was taken.
The EXIF data of an ordinary photo records the capture time, but it usually depends on the device clock. Users can manually change the time, and a compromised operating platform could also provide false time information.
Image | Discover Digital Photography
Therefore, Reference Image also needs a time reference that is independent of the platform clock.
The iPhone periodically obtains signed time credentials from Apple's cryptographic timestamp service, on average about once every 15 minutes globally; after capture is complete, it requests another fresh time credential.
By using two trusted time points, the act of capture is sandwiched between them. Apple cannot prove the exact second the photo was taken, but it can prove that it was produced between these two verified time points.
In other words, Reference Image provides a timestamp that is "relatively imprecise, but absolutely trustworthy." Even if there is no network at the time of capture, the later time credential can be obtained once connectivity is restored.
At this point, the raw pixels captured by the sensor, related metadata, signatures from the sensor and Secure Enclave, the device hardware inventory, and trusted time information are packaged into a Secure Digital Negative, saved on the iPhone in DNG format.
But the negative is not yet the photo we ultimately see.
From RAW to JPEG, a series of image processing steps is still required. If the processing pipeline is tampered with, all the trusted relationships established earlier could collapse at the very end.
Therefore, this negative also needs to be "developed" in a trusted environment. Apple delegates this step to Private Cloud Compute (PCC).
When I choose to generate a Reference Image, PCC first verifies the raw pixels, signatures, certificates, device identity, and time information to confirm that they form a complete and mutually consistent chain of evidence. Only after verification passes does PCC process the RAW data, generate the final JPEG, and download it to the user's device locally.
This entire process feels deeply "old school": it is as if you actually shot a roll of film (corresponding to Apple's Secure Digital Negative), and then took the film to a photo lab (corresponding to Apple's PCC servers) to be developed.
But why should we trust that Apple's cloud service is a "trusted environment"?
Apple's approach is to make the cloud service itself an object that can be verified, rather than relying solely on Apple's one-dimensional endorsement.
The officially deployed software version of PCC is recorded in a cryptographically protected transparency log, and the corresponding binaries are also made available to external security researchers; before uploading the negative, the iPhone also confirms that the target server is running the version recorded in the log.
In other words, Apple is not only trying to prove that the photo's source is trustworthy, but also ensuring that the program responsible for verifying the photo is itself traceable and accountable.
After one final signature from Apple's signing service, we obtain a Reference Image whose provenance can be independently verified. At this point, the entire chain of trust is finally closed.
There is also an interesting detail: PCC's verification does not stop at the photo itself. It once again looks back to the source of the chain of trust — the sensor.
In addition to verifying the cryptographic signature, PCC also runs a neural network to check whether the image has the physical characteristics expected from an Apple camera sensor's raw output, and generates a confidence score.
These scores, along with the photo's unique identifier, sensor ID, and other information, are sent to Apple's revocation service for continuous assessment of the sensor's trust status. If a particular sensor consistently produces anomalous image data, Apple can detect potential security risks and revoke the authentication of photos associated with that sensor when necessary.
In this way, the entire verification process forms a loop: the sensor, through cryptographic signatures, provides a trusted proof of origin for the photo; and the photo itself, in turn, becomes evidence for Apple to check whether the sensor is still trustworthy.
However, when trying to establish a trusted source for photos, there is another easily overlooked challenge.
As mentioned earlier, each camera sensor has its own dedicated key, and Apple can use the device hardware inventory to confirm which iPhone a photo came from.
So, if someone obtains multiple Reference Images, could they use the sensor identity information in them to determine whether they came from the same device, and thereby track the photographer's activities?
For ordinary users, this might just be a privacy concern. But for war correspondents in conflict zones, or photographers investigating sensitive events, exposing the identity of the capture device could mean a more direct risk to personal safety.
Therefore, Apple does not expose the sensor's internal identity credential in the final image. For external verifiers, they can only see the signature Apple gives to the final photo — figuratively speaking, they can only see Apple's stamp saying "captured by a trusted camera."
A very "Apple" kind of product logic
Of course, Reference Image also has its boundaries.
Take the most direct example: I could use an iPhone to photograph a fake AI-generated image displayed on a computer screen. Even if the final photo passes all verification, it can only prove that the camera indeed captured what was on the screen — it cannot prove that the events depicted in the image actually happened.
In other words, Reference Image can provide credible proof of the photo's capture source, but it cannot guarantee that the content shown in the photo is necessarily real.
In its official documentation, Apple also defines a clear and restrained scope of proof for this technology:
A trusted iPhone camera sensor captured this optical data within a verified time range, and this data subsequently underwent verifiable processing.
This also means that, in the AI era, we still cannot simply prove that "the story depicted in the frame is real"; but at least Apple's approach can give a photo an additional authentic record of its own provenance.
Faced with AI images whose origins are increasingly difficult to determine, different companies are trying different solutions.
For a generative AI company, the most natural approach is to leave markers on the content it generates. Google DeepMind's SynthID embeds digital watermarks in AI-generated images that are imperceptible to the human eye but can be recognized by specialized tools.
A traditional camera manufacturer, on the other hand, is more likely to start from the capture device itself. The Leica M11-P mentioned earlier uses in-camera security hardware to generate content credentials for photos taken by the camera.
These approaches have different entry points, but they all start from the links that each company can control, establishing a trusted provenance record for images.
Now, looking back at Apple Reference Image, it becomes clear that this is a distinctly "Apple-style" product capability.
To give a photo a trusted source, Apple has the sensor prove the capture of the original pixels, has the Secure Enclave and device hardware inventory confirm device identity, uses trusted timestamps to determine the time range of capture, and finally leaves verification and final photo generation to Private Cloud Compute.
In this process, every step requires coordination among different components. A key premise that allows Apple to string them together is this: from camera hardware to operating platform to cloud services, all of these links are firmly under Apple's own control.
This reminds me of how Apple created the Dynamic Island.
The punch-hole in the display was originally just a hardware structure. Apple had platform animations, notifications, and app states reorganize around it, and eventually turned it into an interactive interface.
The Dynamic Island's success depends on the coordination between hardware appearance, platform interaction, and app capabilities. Apple considered these capabilities from different technical layers within the same product logic, ultimately forming a complete feature.
Although Reference Image and the Dynamic Island are two completely different capabilities, they share a similar product philosophy behind them.
With the Dynamic Island, the goal was to turn a hardware structure on the display into a natural interactive interface; with Reference Image, the goal is to give a photo a verifiable provenance record from the moment it is captured.
It may just be one of Apple's attempts to respond to the AI era. But starting from a specific problem and then addressing it with the capabilities of an entire product ecosystem — when you think about it, that approach really does seem like Only Apple Can Do.
If you are interested in the technical details behind Apple Reference Image, you can also read the official reference article published by Apple:
https://security.apple.com/blog/apple-reference-image/